Posted by Justin Kubatko on October 16, 2009
Thanks to the alert eyes of one of our readers ("Deepak"), I have corrected a minor flaw in the Win Shares system. For those who don't care, feel free to move on. All others, please continue reading after the jump.
To convert marginal offense/defense into wins I was using league marginal points per win, which is equal to 0.32 times league points per game. However, thanks to Deepak's e-mail, I noticed that teams that played at a below average pace tended to have positive errors (i.e., more wins than Win Shares) and teams that played at an above average pace tended to have negative errors (i.e., fewer wins than Win Shares). The relationship between a team's pace factor and their prediction error was not particularly strong, but it was strong enough that I realized it should not be ignored. The problem was using the unadjusted league average to compute marginal points per win. Teams that play at a faster pace need more marginal points to generate one win, while teams that play at a slower pace need fewer marginal points to generate one win. Marginal points per win are now calculated as follows:
To illustrate this let's look at two teams at the extremes last year. For the the Golden State Warriors, who played at the league's fastest pace, marginal points per win is equal to:
On the other hand, for the Portland Trail Blazers, who played at the league's slowest pace, marginal points per win is equal to:
That's a big difference, and something that should not be ignored. By using the league average rather than the pace-adjusted league average, I was overrating the Warriors players and underrating the Blazers players. In terms of the overall prediction error this did not have a large effect (the overall error did drop, but not by much), but at the player level there are going to be some significant changes. For example, using the old system Andris Biedrins led the Warriors with 5.8 WS, but using the new system he leads them with 5.4 WS. Meanwhile, Brandon Roy, the Blazers leader, goes from 12.8 WS to 13.5 WS.