Reading Room: Physical game, mental game

tl:dr: fWAR for individual pitcher (especially reliever) seasons. don’t take an average of the two. all stats have something wrong with them.

with the samples in question, fWAR is just way better than bWAR. FIP better describes pitcher performance over the year than runs allowed plus a defense adjustment. until you get to the point where you can say for sure that the difference between a component-projected ERA (FIP, xFIP, SIERA, etc.) and actual ERA is real and not random (that’s like 6 full seasons for a starter, iirc), i don’t see why using bWAR makes sense. if you’re looking at team totals or careers, bWAR makes plenty of sense. but for individual player years, strikeouts, walks and ground balls are what matter most. and for relievers, even walks are high variance by comparison.

the other thing is that bWAR uses a much higher replacement level than fWAR. a team full of replacement level players will win ~27% of their games according to fangraphs. per b-r, it’ll win ~32%. so you can’t just add them together and divide by 2 to get an actual average of the two. it’ll end up weighted toward the fangraphs number by some amount.

i checked out the fangraphs and b-r war boards yesterday and in 2010, you could convert bWAR to fWAR by multiplying by ~1.2. that should be close enough going forward to do what you want to do there.*

as far as a better source goes, the ideal WAR would weight by component given a sample size and use a regression. this amounts to, more or less, a continually updating projection as the data from a season comes in. fangraphs has a version for that of ZiPS, but i’ve heard it’s regression factor overweights new data significantly. i’m not sure if they’ve fixed it, but i think i would have heard about that. until then, i use a combo of xFIP and FIP that i kind of mentally weight one way or the other. i’d use just xFIP, but it turns out that our batted ball data isn’t crazy reliable at picking out fly balls versus line drives, so pitchers who actually do a great job at getting the ball on the ground sometimes get shafted because the mix of LDs and FBs is whack.

oy vey. in sum: most of the stats have something wrong with them depending on the sample, the data, etc. this is because we implement them knowing certain things and as studies come out and make new findings, those things are altered somewhat. rather than constantly creating slightly new metrics that could befuddle audiences, they stick with fairly simple stuff that’s good enough for government work.

in this case, i’m running through everything because a)i expect you want to know and b)the small differences are meaningful here.

*and actually, just looking at the the team war leaderboards in each, it looks like they have replacement level as slightly different every year, which i don’t think is ideal. iirc, you’re supposed to set all replacement levels equal to each other for a given system so you can definitively compare them year to year. i don’t think they change enough to really mess with that comparison. it’s just annoying. all the more so cuz it futzes with a post of mine from BtBS i was consulting for my post yesterday.