A comment on my free agent post had enough going on that I thought a quick post was in order. The first part of it was using efficiency difference instead of point difference. To my understanding, this would simply adjust for pace; point difference is per game while efficiency difference is per 100 (or some other number) possessions. Thus a slow team, like the Spurs, may have a bigger efficiency differential than point differential. However, I’m not sure that it’s necessary. The assumption is that, were the Spurs to play faster, they would increase their point differential to come more in line with their efficiency differential; if they are more efficient, and have more possessions per game, they should win by more points per game. I’m not sure that this is true though; it’s possible the Spurs are so efficient because they slow the game down and work a half-court set to get good shots, and that speeding up would actually hurt their efficiency. Does anyone know of any work looking at what happens when slow teams play fast or vice versa?
The second part of it was that were I to use efficiency differential, players could be evaluated by Wins Produced instead of plus/minus (or adjusted plus/minus). I didn’t mean to suggest that the 2 extra points per game I refer to are in a plus/minus framework (I’m actually not a plus/minus fan, in any form); it’s just that my model works on point differential, so the player’s ability would have to be translated to points per game. In reality, the free agent would have to be evaluated relative to the existing team, or what was being given up for him if it were a trade, because the two points per game is relative to what the team did without that player. Or, I could map point differential to wins…
Wins Produced could be used for this purpose. I ran a quick regression of wins on point differential (much like the first step in calculating WP, except that uses efficiency differences), and the equation is wins = 41 + 2.74*differential. The intercept makes sense; a team with a 0 differential wins by as much as it tends to lose by, and so should be average and win half (41 of 82) its games. The slope means that for every point differential you increase, you should win 2.74 additional games. So in this model, the 1996 Bulls would be expected to win 74 games instead of 72, for example. So in my previous post, instead of saying that the free agent would improve your team by 2 points per game, I could equivalently say that he would be worth an additional 5.5 wins produced over whatever your team was giving up (nothing if he’s a free agent, or however many wins an outgoing player produced in a trade).
This would be like trading an average starter (in 2000 minutes, would produce just over 4 wins) for a guy with a WP48 of .232. Obviously that’s the kind of trade you would love to make if you could. You’d think this kind of thing wouldn’t happen, since the trade is so mismatched, but the Nets might have just done even better when they moved Courtney Lee for Troy Murphy in the big four team trade. Sadly, the Nets weren’t starting from a position where this would really help them; on the other hand, Murphy doesn’t kill their cap in the way my previous post assumes. So definitely a great move for New Jersey.
So, overall: I still like my model; it could potentially be improved by using efficiency instead of per-game numbers, but maybe not; and as long as you can tie your player performance measure to point differential or games won, the model should be able to tell you if it’s a worthwhile move or not, whether you’re evaluating a free agent or a trade.