Nathan Sliver, sevenfortyseven.com. Back in 2005, investigative reporter JNR release a revolutionary report going beyond the numbers and providing insight into the locker room dynamics of each team. At the time a blundering yet perceptive GM asked, “Does cohesion affect actual outcomes?” Well, today we feel we have enough data points to take JNR’s beyond-the-numbers insight so far past the numbers that we end up back in the numbers to try to answer the Portland GM’s question.
What does affect “actual outcomes?” There are a few things that we know impact performance: player attributes, matchups, game planning, home vs. away, built-in randomness, etc. Does cohesion belong in that group though? Well, I ran some regressions trying to answer that question and the results were very consistent. First let me say that I only have JNR’s annual data, which means that game-by-game measures such as home/away and matchups are out the window. We also only have a handful of years, which means controlling for GM really kills the degrees of freedom but we know it matters. So that leaves with the following model, with winning percentage as the dependent variable.
- image.png (48.13 KiB) Viewed 1128 times
That 0.033 was very consistent across the many models I tried. This means that for every level increase in cohesion a team can expect to increase their winning percentage by 0.033. So while having good cohesion instead of average may not be the difference between seeds 8 and 1, having average cohesion instead of poor cohesion could be the difference between making the playoffs and scraping the bottom of the lottery barrel (just ask the Blazers). The adjusted R squared of .688 means that these three variables alone explain about 68% of the variation among winning percentages.
So it appears cohesion matters but how, if possible, can a GM manipulate their cohesion? This question proved much trickier to answer. I tried all kinds of complex variables: number of returning players, total years with the team of returning players, PER of returning players, average winning percentage over the past 2,3,5 years. After much trial and error, I only found two variables that were fairly consistently significant. One is simple, the previous season’s winning percentage. The other was very complicated. It is a measure of my own device. It was a measure of how players on the team the previous year performed relative to the predictions of my player performance model. Over performing (even if that wasn’t great performance) was associated with better cohesion (see below). However this model only explained about 20% of the variation of cohesion.
- image (1).png (21.34 KiB) Viewed 1128 times
Then I had a thought. What if my model was tautological? What if the reason the team underperformed the previous year was because they have good cohesion. So I tried the same model but included the team’s previous cohesion. That one variable sucked up all the variation and explained 44% of the variation all by itself. What does this mean? It means cohesion is sticky, it doesn’t change too much year to year. Very few teams can make a two category jump. So what should a GM do if his team is stuck with bad or poor cohesion? I think the best model is the Heat. The Heat missed the playoffs 6 seasons in a row and had bad cohesion before suddenly jumping to average and then good. What did they do? As far as I can tell, they built a strong foundation of young prospects, which meant losing a lot. Then, when they knew they had a winning core, they pulled the trigger on deals that made the team too good to fail.
- image (2).png (22.73 KiB) Viewed 1128 times
