SABR 301- Win Probability Added (June 26, 2003)
This was my first stab at it.
--posted by TangoTiger at 09:42 AM EDT
Posted 12:47 p.m.,
June 30, 2003
(#1) -
tribefan08
Tango,
When Drinen does his WPA he uses a simulator to find the individual park win expectancies. When you did your quick attempt at WPA did you use park factors or did you just use the Win Expectancy table that is listed on your site?
I am trying to compile a list of the best relievers in the game using win expectancy. What do you suggest would be the best way to go about this. I know you suggested LI times ARP at one point and then retracted that statement. Is there a way to calculate ARP including inning/score states? Or is WPA the best measure by itself? If you do think I need the park adjusted win expectancies, how would I get these if I don't have a simulator? Thanks Tango. P.S. this site is dead when you aren't around.....
Posted 8:37 a.m.,
July 1, 2003
(#2) -
Sky
Tango, in the Win Shares pdf you and Rob Wood put together, the main conclusion was that Win Shares "fails" at a basic theoretical level in its attempt to assign absolute win shares to specific players since it can result in players getting credit for more win shares than game shares. Is this a problem inherent to James' Win Shares method specifically, or any attempt to assign absolute wins? It seems like WPA would run into the same problem (not that you're trying to give credit for absolute wins, just marginal wins above average).
Wouldn't it make more sense, when trying to assign absolute wins, to compare every player to the worst possible players (100% SO for hitters). Only a team of perfectly bad hitters could manage to lose every single game. Replacement players win their fair share and should still get credit for some absolute wins, if not many. And on the flip side, the only team wouldn't lose any game would be a team of perfectly good hitters (1.000 OBP) since even a team scoring 20 runs per game would still lose occassionally.
What if, to measure Bonds' absolute wins, one computed the expected wins with him in the lineup (actual, pythagorean, whatever), and then subsituting in a .000 OPS guy in his spot. That would give his absolute wins, no? The other end's a little harder, since as your near infinite run scoring, all positive hitting outcomes converge towards equal value, but in the MLB environment, they definitely aren't. Would you say a perfect hitter gets all walk? All doubles? All homeruns? There's a big difference.
Posted 11:06 a.m.,
July 1, 2003
(#3) -
Rally Monkey
"Wouldn't it make more sense, when trying to assign absolute wins, to compare every player to the worst possible players (100% SO for hitters). Only a team of perfectly bad hitters could manage to lose every single game."
For Win Shares, James is not comparing batters to a replacement level, and also not to players who strike out 100%. He's somewhere in between, assuming that a team that scores 50% of average and gives up 150% of average will be just bad enough to lose all the time.
When you talk about perfectly bad hitters, then we have to get into negative win shares. Lets say you add Bonds to this team. He gets walked every time, and should get credit for this. After he's walked, the next 3 guys strike out and Bonds is stranded. The team still scores zero runs, so the whiffers have to have negative win shares to offset Bonds' positive performance.
"What if, to measure Bonds' absolute wins, one computed the expected wins with him in the lineup (actual, pythagorean, whatever), and then subsituting in a .000 OPS guy in his spot. That would give his absolute wins, no? "
That would give you the total of his wins and the negative wins a .000 hitter would have.
At the other end, it really doesn't matter if the hitter gets all walks, all homers, etc. Its still perfect. This team would never win a game, though, because the game (or inning) would never end. Somebody had to point this out to the Red Sox last Friday.
Posted 7:12 a.m.,
July 2, 2003
(#4) -
tangotiger
I've only got a sec to reply to the first post.
No, I do not yet adjust for park or league or any other environmental condition (opposing pitcher or batter, etc). But I should.
To measure reliever's past effectiveness, you need some form of WPA, meaning inning/score/base/out at least, plus park/league preferably, and batter/fielders, too.
Wolverton does only base/out, which is also pretty good. For the moment, your best bet is to stick with Wolverton.
As for using LI, the thing is that that assumes that the pitcher is equally effective in all base/out, which is one reason you can't simply multiply it to Wolverton's. The other is that Wolverton already "LI's" the base/out, in effect. Therefore, I'd have to give you an LI based only on inning/score, for you to multiply by Wolverton's number.
You other option is to calculate a reliever simply by using "peripheral ERA", and then multiply by LI. That might get you part way there.