from Amazon
Tuesday, February 02, 2010
Various implementations of WAR; talking WPA/LI
The framework of Wins Above Replacement (WAR) was developed in this blog over a period of months or years.
Nonpitchers:
- Offense relative to average
- Fielding relative to average
- Positional adjustment
- Common replacement level adjustment
You add those up, and multiply by expected/deserved playing time.
Pitchers
- Pitching relative to average
- Role adjustment (starter / reliever)
- Leverage adjustment (reliever)
- Common replacement Level
You add those up, and multiply by expected/deserved playing time.
That’s the framework of WAR. There are different implementations of this framework, as Fangraphs has it (fWAR) and Rally’s Baseball Projection has it (rWAR). They each make some decision as to how to count each component. The great thing about the framework is how you can slide one thing in or out without affecting anything else. Prefer UZR to TZ, but like everything else Rally did? No problem, slide one out, slide the other one in.
Jeremy doesn’t like some of the choices made by some of these implementations.
My main philosophical problem with Fangraphs’ WAR (fWAR) is that relievers are given extra value for having pitched in high-leverage situations. Personally, I don’t understand why we use a pitcher’s actual leverage index and chain from there. Why not just start and end with the deserved leverage index?
In terms of forecasting, I definitely go with deserved LI. But, in terms of accounting for the past season, I don’t know that I would do that. After all, imagine if Mariano Rivera was used in mop-up duty all season. His actual win impact was muted. Imagine Ozzie Smith at DH or 1B. Imagine Adam Dunn in CF. So, in terms of accounting for actual wins and losses, the actual usage is what we care about, and not the optimal / deserved usage.
But, that’s fine. We can disagree on it. Jeremy can have his jgWAR if he likes. As long as we adhere to the basics of the framework, then 90% of the disagreement goes away. Now the conversation moves to the periphery. And that’s a good thing. It’s no longer an RC v BsR debate, but a debate as to the specific component of the “B” variable in BsR. That’s good.
Now, Jeremy also brings up WPA/LI for pitchers. In this particular case, there’s a bit of a problem. WPA/LI is great for hitters because it neutralizes all the PA so that all the PA are equally impactful. It simply recalibrates each of the components. Basically, it’s like having a game-state specific version of wOBA. Sometimes the coefficient for the HR is 1.5 and walk is 0.9, or sometimes the coefficient for the HR is 3.0 and the walk is 0.4. There’s always some sort of recalibration based on how much impact the walk and HR will have on that particular game state. But, for pitchers, it’s not so easy. Prime-Pedro has many fewer men on base specifically because Prime-Pedro is pitching. So, we don’t want to neutralize each PA so that they count equally and then add them up in a linear fashion, which is what WPA/LI does. WPA/LI is the first step, but then it needs to be BaseRuns-ized in order to get it into the right scale.
How big a deal is this? Interestingly not big at all. When I look at the big 4 of our generation, their career WPA/LI and career WPA has almost no difference. In any case, to the extent that we have an issue, the bias would work the same for all the same-quality pitchers. Basically, this is one of those things that is not worth worrying about, other than if you are really someone who enjoys digging into an issue like this. So, yeah, WPA/LI (for pitchers) gets you at least 95% of the way there, if not 99%.
I love WPA/LI because it balances every PA to be equal to any other PA, while also adjusting for the particular vagaries of the situation. A runner on 3B and less than 2 outs and down by 1 run is not the same as with 2 outs and up by 5. WPA/LI handles it properly, and is the only metric to do so, other than its cousin WPA.