Chemistry and synergies in the NBA

[eugeneshen]

What if you have a 20%-winning team, and you add David Robinson and Tim Duncan?
The Spurs did that and became a 70% winning team.
Is there any way to predict these things?

David Robinson, yes, because he had already played in the league. Tim Duncan, doubtful, since he was a rookie. But yes, if the data is available, you could theoretically use the framework to forecast wins and losses, although there are a lot of complications: forecasting minutes played, the effect of aging, etc.

[eugeneshen]

Do you have any opinion / view based on training & experience about the value of computing blended metrics for decision-making instead of going with a single metric? Say SkillPM and RAPM or a more involved blend with additional metrics? Is this potentially or likely an optimal strategy or mainly just risk-adverse (or do these coincide?)?

I think it makes sense to look at a variety of metrics, quantitative and qualitative, for decision-making. More information is always better than less information.

[pmaymin]

Any chance you might share the full SPMs for the entire league to allow both full appreciation of your work but also aid further comparative work?

Rich Cho is said to have a tool that gives every player a specific single number value for trade considerations. Your trade consideration approach would be quite different by emphasizing top lineup specific values in the new contexts (starting lineup or maybe extended to the basic rotation set of lineups).

Yes, we are working on a website that would allow you to explore the impact of trades, and check any lineup vs. any other lineup. Will let you know when it is ready!

[pmaymin]

Still I am somewhat intrigued by the BOS-MIN Rondo-Jefferson trade mentioned. Would that hypothetically have been for the 2010-11 season or 2009-10?

The one intrigued us too! BTW -- everything is for the four seasons of 2006-2007 through 2009-2010, all pooled together.

The best way I understand this is that actually Rondo comes up in a lot of mutually beneficial trades, and often for the kinds of people we'd call scorers or shooters, like Michael Redd or Jamal Crawford or Ben Gordon or Carmelo Anthony. My best guess at this point is that it is similarly Big Al's offensive prowess that would fit in better with the rest of the relatively defensive minded C's.

Perhaps (speculating wildly) this is some of the reasoning that went into the acquisition of relatively offensive-minded Jeff Green for defensive stalwart Kendrick Perkins. Either way, I'd bet with Perk gone, a lot of the Rondo-for-offense synergy trades would no longer be mutually beneficial.

What were the best mutually beneficial trades identified (if you are willing to say) and how big gain could both sides yield in the most beneficial such trades? Have any of these 222 happened yet, in any altered overall form?

Surprisingly, Rondo for Nate Robinson would have been mutually beneficial, but Boston actually got Nate without giving up Rondo.

[kggk]

I have a question about synergies of abilities with themselves. For low probability events, like steals, is it an assumption of your model that they have positive synergy? In other words, is this just a consequence of the form of your model, and the second derivative of the cdf of the normal distribution? Or could steals turn out to have negative synergy in your model if the data was different?

Thanks

[Crow]

Thanks very much Eugene and Philip for the multiple replies to my questions and the questions of others. I really appreciate it.

I look forward to seeing your site.

[eugeneshen]

I have a question about synergies of abilities with themselves. For low probability events, like steals, is it an assumption of your model that they have positive synergy? In other words, is this just a consequence of the form of your model, and the second derivative of the cdf of the normal distribution? Or could steals turn out to have negative synergy in your model if the data was different?

That's right. For low probability events like steals, positive synergies come from the second derivative of the normal cdf. We believe this specification is realistic. If your opponent has a 1% chance of creating a steal, then introducing another great ballhandler does not help much. Or, if you have a 99% chance of grabbing a defensive rebound, adding another rebounder make no sense. Steals could have negative synergies if they become commonplace, for example if Rondo plays against a bunch of high schoolers.

[Crow]

Crow wrote:
Is it true that total synergy of a lineup would be the sum of the overall net synergies of 10 player pairs and 21 skill synergies per pair for a total of 210 synergies?

reply: No, the total synergy of a lineup is calculated as PORP(p1,p2,p3,p4,p5) - PORP(p1,rp,rp,rp,rp) - PORP(p2,rp,rp,rp,rp) - PORP(p3,rp,rp,rp,rp)- PORP(p4,rp,rp,rp,rp)- PORP(p5,rp,rp,rp,rp), where PORP is point differential vs. a team of replacement players, p1-p5 are players, and rp is a replacement player.


Ok the total synergy of a lineup is calculated as described, but in reality it would still ultimately be based on the net total synergy arising from 210 skill synergies.

If RAPM for player pairs were taken down to the 4 Factors (or if SkillPM was for its 3 skills) then one could look at specific synergies between teammates and between player and opponent and understand better what goes into the total net lineup synergy.

[Crow]

Trying to boil down or at least restate a thick / probably hard to follow previous passage a bit that is still, in my mind, on the table for comment:

The article's results for synergy for starting lineups show no more than a 10% change from sum of separate player ratings to combined lineup rating. So essentially is the article saying player synergies are pretty low and not that important? Or less sweeping, that experienced net synergies in actual lineups, given the way they tend to be constructed, are almost always below 1 point? What was the largest net synergy seen in a lineup above a trivial minutes of use?

My previous comparison of sum of separate player RAPM ratings and combined lineup RAPM ratings for most used lineups caused me to lean toward thinking that specific lineup constructions and their specific synergies could in some cases (including for some big minute lineups) appear very important (at for a single season; I haven't checked enough multi-season to be confident to make that claim) if the variance between expected lineup performance and actual performance were considered real and not just or mainly noise.

Any further discussion on this issue?

[eugeneshen]

What is the average lineup impact rating of the average lineup played in the league? Is it close to 12-15?

Great guess! The average PORP across our entire data set is 2.82pts, or 14.1 per lineup.

[eugeneshen]

The article's results for synergy for starting lineups show no more than a 10% change from sum of separate player ratings to combined lineup rating. So essentially is the article saying player synergies are pretty low and not that important? Or less sweeping, that experienced net synergies in actual lineups, given the way they tend to be constructed, are almost always below 1 point? What was the largest net synergy seen in a lineup above a trivial minutes of use?

Synergies range from -0.8pts to 1.2pts. We believe that these numbers are significant. Using the Pythagorean formula, 1-2 extra points would mean 3-6 extra wins for a .500 team. Six wins could be the difference of not making the playoffs versus being the #4 seed!

[mtamada]

What if you have a 20%-winning team, and you add David Robinson and Tim Duncan?
The Spurs did that and became a 70% winning team.
Is there any way to predict these things?

Well if you want to play that game: the Spurs in 1997 won 24% of their games.

What happens if they add Robinson and Duncan? In 1998 they won 68% of their games.

What happens if they add just Robinson? That's what they had in 1996 -- and they won 72% of their games.

What happens if they add just Duncan? That's what they've had since Robinson retired -- and they've continued to win 61%-74% of their games.

That doesn't mean that a team wouldn't love to have both Robinson and Duncan. But it does mean that the second Hall of Fame big man adds a lot fewer victories than the first.

It also helped that Duncan though a low post player has been as much a power forward as he has been a center. And that Robinson was a multi-skilled player able to contribute even when he wasn't posted down low. So they could play together effectively. A combination of say Elvin Hayes and Artis Gilmore would be more negatively synergestic. Whereas Hayes and Unseld was a dynamite combination.

[Crow]

Thanks for the follow-up on the average lineup impact rating of the average lineup played in the league.



"Using the Pythagorean formula, 1-2 extra points would mean 3-6 extra wins for a .500 team."

Only the extreme starting lineups are near 1 point of synergy up or down (per game or per 100 possessions?) and only a few starting lineups play a full 1/3rd of the game.

Is it possible to compute the average absolute lineup synergy for all lineups, minutes weighted?

The simple average absolute lineup synergy for all starting lineups would appear to be just over 0.5 pts per game or 100 possessions. Actually playing just 10-16 minutes per game, that would suggest an average absolute synergy impact from the starting lineups of 0.1 to 0.2 pts per game up or down or perhaps a 0.3 to 0.6 win impact per season up or down? If this calculation is correct.

[eugeneshen]

Only the extreme starting lineups are near 1 point of synergy up or down (per game or per 100 possessions?) and only a few starting lineups play a full 1/3rd of the game. Is it possible to compute the average absolute lineup synergy for all lineups, minutes weighted? The simple average absolute lineup synergy for all starting lineups would appear to be just over 0.5 pts per game or 100 possessions. Actually playing just 10-16 minutes per game, that would suggest an average absolute synergy impact from the starting lineups of 0.1 to 0.2 pts per game up or down or perhaps a 0.3 to 0.6 win impact per season up or down? If this calculation is correct.

You bring up a good point but forget the fact that the backups also have potential synergies. Assume for a moment that the starters and backups play equal minutes, and that starter synergies and backup synergies are independent. Then the variance of the total synergies would be 1/2 starter variance + 1/2 backup variance. Thus, the total standard deviation should be 1/sqrt(2) of the starters. That's probably a lower bound. So the range would be at least 2-4 wins per year, which is significant. Does that make sense?

[Crow]

I didn't forget that units beyond the starters could have synergies, I just didn't assume it in what I selected to say because I wanted to give the starter's synergy a proper minute weighting and I don't know the size or sign of the synergies for non-starting units. At some lower level of starters in the unit, the sign of the lineup synergy for a "bench unit" might tend to become negative on average and partially cancel the synergy of the starting unit and that would again lower the win impact.

I am not asserting what the win impact of all lineup synergies is precisely, just suggesting that they may be lower or considerably lower than was first mentioned. The win impact is certainly not of the scale of a star's or regular starter's impact and it appears they are probably less than that of many top subs. A 4 win impact would be 10% of the average team's outcome, a 2 win impact, 5%.

To get to a 4 win impact you'd need an average 1.3 point impact across all lineups, or in other words you'd need the average synergy impact to be higher than any starting unit impact in the league measured. I don't think that is likely to happen on average or even occasionally. Even a 2 win impact would require that the average synergy impact of all lineups for a team be above 0.6 when the average absolute impact of starting units was just about 0.5. Unless all the lineups had a synergy of the same sign that would work to reduce the distance the average got from 0. Even a 2 win impact is probably going to be rare, if it happens, based on the numbers I've seen.