APBRmetrics

Crow, maybe yet another way to measure the importance of synergy is to consider a team that is drafting or trading for a player and has a choice between two players of equal total talent but different skillsets. Let's say the good one adds 0.5 points per game of positive synergy and the other subtracts 0.5 point per game of negative synergy. Then the wrong decision costs about 3 games per season.

If each win is worth about $2mm, that's $6mm per year of free money the team is losing by getting a bad chemistry fit. That's enough to afford an extra free agent per year.

And of course a decision like that can have ramifications for more than just one regular season, and for playoffs too, so it is even more important.

Ultimately of course you need talent! But at the margin, making the wrong decision on synergy could end up hurting you.

Sound reasonable?

BTW thanks again for all your comments and feedback. This is fun stuff!

Another perspective: what's the average margin of victory in NBA games? Is it about 4 points? Then a good-synergy vs. bad-synergy decision that nets you an additional 1 point per game could be quite important.

Yes the dialog is a good thing. And synergy is important, every little bit can matter. Up or down part of or a full point is significant especially if it is up vs down. Synergies may be most important for teams near making the playoffs or near making the next higher level.

While I was discussing the indicated synergy levels and asserting they were pretty small, I still wonder if they might actually be bigger than indicated.

Nelson-Bogans-Turkoglu-Lewis-Howard for the Magic in 2008-9 were +20.6 on raw +/- and estimated at +22.5 on traditional Adjusted +/- at basketball value in 276 minutes of use. The standard error was reported at 7.5.
The sum of the 1 year player APMs was about +18. I guess the overperformance of raw +/- compared to the sum of the player APMs was within expected variance for the sample size.

Williams-Gibson-James-Varejao-Ilgauskas for the Cavs in 2008-9 were +34.7 on raw +/- and estimated at +34.2 on traditional Adjusted +/- at basketball value in 66 minutes of use. The standard error was reported at 15.5. The sum of the 1 year player APMs was about +13.3. The overperformance of raw +/- compared to the sum of the player APMs was much bigger than in this case and larger than 1 standard error even inflated for the small sample size. Maybe it is mostly noise or maybe it is not. Is there anything more that can be fairly said?

Paul-Butler-Stojakovic-West-Chandler for the Hornets in 2008-9 were +7.8 on raw +/- and estimated at +2.5 on traditional Adjusted +/- at basketball value in 424 minutes of use. The standard error was reported at 6.2. The sum of the 1 year player APMs was about +6.3. The overperformance of raw +/- compared to the sum of the player APMs was small in this case and the underperformance on lineup APM was by a bit over half a standard error.

Kidd-Terry-Howard-Nowitski-Dampier for the Mavs in 2008-9 were +11.4 on raw +/- and estimated at +13.5 on traditional Adjusted +/- at basketball value in 252 minutes of use. The standard error was reported at 7.7. The sum of the 1 year player APMs was about +7.5. The overperformance of raw +/- compared to the sum of the player APMs was moderate in this case and the overperformance on lineup APM was by nearly a standard error.

On RAPM for the 3 of these lineups with enough minutes to see the lineup values (from multi-season RAPM) the over and under performances were by 2, 2 and 7 points. No standard errors were available. Again the difference between lineup performance and sum of individuals (the apparent "synergy") is larger than the estimated true synergy by SkillPM.

Maybe what I am seeing from these checks is mostly noise. But what do you do then? If you are making a decision in a game would you follow small sample noise (and guess / hope it might continue and do so until it doesn't) or ignore it? Which would more likely do better in that imperfect context than the other option? The specifics matter, the size of the overperformance relative to the standard error but I'd think it would be better to follow some of these overperformances / possible positive synergies than not to.

More testing of these lineups would reduce the standard errors and improve the ability to spot truly strong or weak synergies. If playoff teams tend to use say 10-15 lineups for almost all the playoff minutes, does it really make sense to use a lot of the regular season time available on hundreds of small minute lineups? There is a balancing act between matching up or freewheeling in the process of winning regular season games and more intensive testing to evaluate most used lineups and find the best but if you are a good team and really focused on contending and winning the title I'd put more priority on testing most relied upon lineups than is typical in the NBA today.

Crow wrote:Maybe what I am seeing from these checks is mostly noise. But what do you do then?

Great point, Crow. Yes, I agree that you can use Lineup APM minus individual APM to measure synergy. But as you noted, there often is not enough data points to draw a conclusion, because the number of lineup combinations is too high. Our paper's innovation is that we are able to predict synergies without encountering this exact small sample problem.

Yes, your paper provides a clearer, simpler to interpret message on positive and negative synergies and I appreciate having that guidance for some lineups now and hopefully all later.

I'll probably still look at the data from APM or RAPM player to lineup comparisons but I will temper my enthusiasm for the strong and weak by this method quite a bit, knowing the results from your method.