Community - FanGraphs Baseball


RSS feed for comments on this post.

  1. This is super cool, but Voros McCracken never claimed that pitchers have no control over what happens to batted balls, just that it was not something that was consistent enough to be predictive, and therefore isn’t useful for predicting which pitchers will do well in the future. Very different claim.

    Comment by Eminor3rd — October 30, 2012 @ 4:05 pm

  2. Thank you!

    In my defense, I did say that he said they have “practically” no control, and by that I meant what you said. I think, hopefully, I’ve taken a step towards demonstrating there actually is a practical way to make distinctions between pitchers (especially in the long run) that may have eluded McCracken.

    Comment by Steve S — October 30, 2012 @ 4:16 pm

  3. Great stuff, Steve. One thought: now, I know w/ values per-pitch type are generally considered insignificant due to the BAPIP on them, but I have found in my own research to show significant results on what is lucky in one particular combination: high FA and high CH together appear as “lucky” when you look at where those pitchers rank in terms of HR/FB, BABIP, LD% and LOB%. So while one can claim that those values are luck based, the fact that far, far more percentage of “plus” FA plus “plus” CH” are seen as lucky, it happens at a far more significant rate than any other combination of “plus” pitches, year after year. Thoughts?

    Comment by Will H. — October 30, 2012 @ 6:45 pm

  4. Thanks Will. Sorry, I’m a little confused about what you’re saying; are you referring to velocity, movement, location, frequency, or some combination thereof regarding fastballs and changeups?

    I guess since 4-seam fastballs and changeups are the first and third most common pitches, that could help explain why they’re so significant, though (about 45% and 11%, respectively, with sliders in 2nd at about 15%). But I definitely see the value in being able to throw off a hitter’s timing the way a good fastball/changeup combination can, which could lead to less contact on the “sweet spot,” swinging underneath fastballs (when they drop less than the hitter expects), and maybe more defensive swings.

    Comment by Steve S — October 30, 2012 @ 7:31 pm

  5. Some very interesting stuff here. I like it. With that said, both factors (LD and IFFB) seem analogous skills of preventing batters from squaring up on the ball when they make contact. IFFB would be getting too far underneath the ball, while a low LD percentage would be the ability to keep it off the meat of the bat on either side.

    It makes me think that an important piece of information is the distribution of angles where the ball hits the bat for that pitcher. Balls struck above a certain angle are likely to be outs (regardless of if they are IF or not, they’re likely to be caught before dropping). Balls below a certain angle will be ground balls. Not sure if anything captures this kind of information, but an ability to throw pitches that influence this distribution should result in these kind of effects.

    Comment by B N — October 30, 2012 @ 9:52 pm

  6. Thanks B N. Yeah, it would be really nice to have access to Hit F/X data for those reasons, but it’s not publically available, as far as I know.

    I did find it odd that there’s very little connection between popups and line drives allowed — the correlation was 0.066, which means pitchers who generate more popups actually tend to allow more line drives (but the two have almost nothing to do with each other). I would have expected pitchers who got more popups to allow fewer liners, by virtue of being harder to hit in general. Apparently it takes two different types of approaches that are not very aligned.

    I thought this was a very cool chart that demonstrates some of what you’re talking about, though.

    Comment by Steve S — October 31, 2012 @ 3:23 am

  7. Using data from BIS, league average LD% changes a lot from year to year. Furthermore, the changes have not been well correlated at all with BABIP, giving me the impression that it’s a measurement difference, not a real change. I think that’s going to make it difficult to achieve a very high level of predictability to take batted ball data from year 0 to predict BABIP in year 1.

    Comment by Detroit Michael — November 1, 2012 @ 7:33 pm

  8. Yeah, I think you’re probably right, Michael. What appears a line drive to one stat recorder might not to another. Hit F/X data should be more precise. I think we’ll find from that data that the some pitchers really do get hit harder than others, in the long run, and that a good deal of it is their own fault.

    As for using year 0’s batted ball data to predict year 1 BABIP… you’re right. With this formula, the correlation was only 0.166. I have another one that gets 0.278, though, thanks to its greater emphasis on popups (far more consistent year-to-year). I’m going to work on something that does a better job of predicting future years than that for the next article.

    Also, using years 0, 1, and 2 to predict year 3 makes a huge improvement over just using 1 year of data, getting correlations over 0.4 for 2012, despite it being an unusual year for LD%. I’m working on looking at other years for the article.

    Comment by Steve S — November 1, 2012 @ 8:34 pm

  9. Steve,

    This is a great article, and I find it fascinating to see the correlations that all of these metrics have to BABIP. I think that your xBABIP approximation should be very useful and is better than any other approximation I have seen. I really liked the idea to use FB%*IFFB%. This value seems more intuitive to include in the regression than what people have previously used. The only concern that I have is that I am not sure that LD % and FB%*IFFB% are independent. I think this because if LD % increases, there is inherently less probability mass available for the FB%*IFFB% to occupy.

    This would be easy for you to check by taking the correlation between the two inputs. If a strong correlation does exist, you could improve your approximation by using principal component analysis to orthagonalize your inputs. Then, you can rerun the regression with independent inputs.

    Comment by Kevin T — November 6, 2012 @ 12:24 am

  10. Thanks Kevin,

    Well, the correlation between line drive percentage and FB%*IFFB% is 0.066 over the 2002-2012 data set (I called FB%*IFFB% “popups” in my reply to B N earlier). I definitely get why you would think they’d be correlated, though (I thought so too, early in the analyses). I should have stated in the article that there was basically no correlation there. But, just assume that if you don’t see any of the batted ball or plate discipline numbers in those first three tables, it’s because the correlation was weaker than the others on the list (and therefore pretty insignificant).

    My explanation for the lack of a connection between the two factors is that popup-inducing pitchers tend to be aggressive, with “rising” pitches, which are both apparently detrimental to preventing liners; this is counterbalanced, for one thing, by a possible unifying link of weaker contact in general (with HR/FB as a mediocre stan-in for the speed off the bat).

    Comment by Steve S — November 6, 2012 @ 2:55 am

  11. I like your article, and it was referenced and used in this article at The Crawfish Boxes:

    Comment by CJ in Austin, TX — November 6, 2012 @ 10:35 am

  12. Awesome, thanks CJ. Great article over there.

    If people are curious how I came up with that formula, yeah, it definitely wasn’t arbitrary — I had my computer come up with the most accurate (in terms of both correlation and RMSE) numbers it could. I started out with a lot more factors, including GB%, FB%, and HR/FB, but discovered they didn’t make a difference, really, as I started whittling them down. The numbers for the two remaining factors were 15 digits long, but they were pretty close to 0.4 and -0.6, so I just rounded them with a pretty minimal cost to correlation coefficient (I figure there’s no point in trying to be very precise when it comes to BABIP).

    Comment by Steve S — November 6, 2012 @ 2:07 pm

  13. This is an interesting article, I love how you broke down the cause’s of Line drives, and IFFB’s. I’ve done a lot of research in this area as well, and developed a similar conclusion: ( However, since batted ball data (particularly LD%, and IFFB%) vary so much themselves (as you pointed out), I’ve found the equation to not be very useful. While it (my equation, as well as yours) do a good job of dissecting what makes a pitchers current year BABIP. It doesn’t appear to do a good job of predicting future BABIP. And it’s easy to see why, as LD% and IFFB% can swing so much from year to year (as you pointed out, even moreso then BABIP itself).

    On that front, for my projection system this past year, I used an equation that incorporates K% (as you pointed out, strikeout pitchers tend to have higher BABIPS), GB% (flyball pitchers have lower babips), an adjustment for park factors, and an adjustment for team defense (specifically I used last years UZR, then did a kind of manual guess of next years team UZR, by moving around numbers for some of the best defensive players who moved teams). The result, it correlated, and RMSE’d better then previous year pitcher BABIP, and Bill James projection system (the only one I benchmarked against). With Bill james at a .034 and me at a .029 RSME it’s about a .005 improvement. Unfortunately, with park factors, and team defense components, my equation needs to be manually built every year, and I’m not sure if the results were good enough to warrant doing it again for another year.

    Comment by slash12 — November 12, 2012 @ 9:24 am

  14. Hey slash, thanks for reading. I definitely remember reading about some of your hitter BABIP work on Fangraphs before, when I was working on some of my own.

    Did you misspeak on the K% vs. BABIP relationship? I actually found that high strikeout pitchers tend to have lower BABIPs.

    Can you tell me more about your latest projection system? What was the sample you used to get that .029 RMSE — what was the minimum innings pitched, and was the projection made from only one year of data, or several? I’ve been working on my own, and I want to know if I have anything worthwhile here.

    Also, my projection system doesn’t include park effects or defense factors, due to the noise I think they probably add to it all (and the complexity), but I was curious about how much of a difference they made for you.

    Comment by Steve S — November 13, 2012 @ 3:52 pm

  15. yes, I misspoke.

    The sample is 113 pitchers in 2012, I projected them at the beginning of 2012, recorded bill james projection as well, and then compared the BABIP results at the end of the year, to find that Bill James had a .163 Correlation, and .034 RMSE, I had a .396 Correlation, and .029 RMSE.

    I scrapped the batted ball data for the most part, my BABIP projections were build like this:
    1) I build a park factor for BABIP, using historic BABIP data by park (3 years), then I halved it (since they only play half their games at home)
    2) I build a defense factor for BABIP, I did a multi-year regression of team UZR as it related to team BABIP. Then I did team UZR projections for 2012, using a very rough system of using 2011’s team UZR, then adjusting them manually for players who were removed from, or added to teams, who had a defensive impact of one extreme or another (the best, or worst defenders). Again, I halved the impact, this time because I just didn’t trust my rough estimations of projecting team UZR
    3) I build a regression of 3 years of previous data (2009-2011) to see how GB% affects BABIP (it increases it)
    4) I build a regression of 3 years of previous data to see how K% affects BABIP (it decreases it)
    5) I combined all these factors together to come up with my projection.

    In the end, I did ERA Projections as well, based on this projected BABIP (and my projections for other stats), and it performed better then Previous year’s SIERRA, Bill James, and ZIPS. But it didn’t perform so much better that I’m going to do it again this year.

    Pre-season last year I did some comparisons to find that my batted ball data based equations (like yours, based around LD%, and IFFB% mostly) just were not doing a good job predicting subsequent season BABIP (previous year’s BABIP/Multi-year BABIP was doing a better job predicting). This park factor, K%, GB%, Defense factor method does a better job (and would do an even better one, with real UZR projections).

    Comment by slash12 — November 15, 2012 @ 9:57 am

  16. One thing I didn’t mention…Part of the reason any of my projections outperformed Bill James, or Zips, can be chalked up to the fact that I knew about, and accounted for team/league moves, that they were not accounting for (because their projections were built before some players even switched teams). In fact, that may be the entire reason my projections were better, I’m just not sure. I haven’t done the work of eliminating these players and seeing how things stack up, partially because doing so would make a sample that’s already probably too small, even smaller.

    Comment by slash12 — November 15, 2012 @ 10:01 am

  17. Great article. Just wondering, how did you calculate the correlation of Pitch F/X movement to BABIP? I’ve been looking for a place with raw F/X data for a while.

    Comment by Sabermetric Solutions — November 19, 2012 @ 7:11 pm

  18. OK, thanks slash. Well, I think I must have taken a different approach from yours, because I’ve seen more success with the batted ball data than you seem to have. I think I’d better write an article about it after all.

    Comment by Steve Staude. — November 19, 2012 @ 11:51 pm

  19. Thanks, Sabermetric Solutions.

    I got all the data right here at FanGraphs. Here’s a link to the vertical movement data.

    Comment by Steve Staude. — November 19, 2012 @ 11:57 pm

  20. Question for Steve and/or Slash. I read the article by Slash in the link above, and he wrote this:

    “Fantasy baseball is one example of a case where FIP doesn’t necessarily do us a lot of good. In this case we’d rather get an idea of what their real ERA is going to look like.”

    Is this because FIP doesn’t take into account the park and team factor? Does FIP only show what the ERA would be if pitching for a neutral team at a neutral park?

    I think I read that Fangraphs FIP- does account for park factors, is that correct?

    Comment by Larry — November 21, 2012 @ 5:26 pm

  21. Hi Larry, I just crunched some numbers, to get an idea of the effectiveness of predicting ERA based on certain stats. But first, to answer your question:
    The formula for FIP is:
    FIP = ((13*HR)+(3*(BB+HBP))-(2*K))/IP + constant
    The constants vary by year (listed here, as cFIP: )

    So, parks have some effect on FIP, via HRs (and, arguably Ks, in the case of breaking balls that break less in Denver). Probably thanks mainly to the DH, the leagues should definitely have an effect, though.

    I wasn’t able to locate the formula for FIP-, but what it does is attempt to park- and league-adjust the pitcher’s FIP. So it’s FIP- that [purportedly] shows what the pitcher’s ERA would be in a neutral park and a neutral league.

    Anyway, here’s how well some of those types of stats in one year correlate a pitcher’s ERA in the next year:

    FIP: 0.432
    tERA: 0.430
    SIERA: 0.407
    FIP-: 0.393
    xFIP: 0.392
    xFIP-: 0.385
    ERA: 0.338
    AVG: 0.308
    ERA-: 0.305
    WHIP: 0.294

    These were the averages for my 2002-2012 sample. Here are the mean absolute errors from the next year’s ERA for the relevant stats:

    FIP: 0.631
    xFIP: 0.634
    SIERA: 0.639
    tERA: 0.712
    ERA: 0.720

    So, FIP, despite being one of the simplest of those methods, is the best, it seems to me, for predicting the next year’s performance. I’m sure I’m leaving out plenty of other methods, but those were the ones I got off of Fangraphs.

    Comment by Steve Staude. — November 22, 2012 @ 12:13 am

  22. Steve,

    Thanks. Let me ask another way…not sure if you can answer this, but if not, maybe someone else can:

    Suppose I have the following data for 2012 pitchers:

    Pitcher Team Actual ERA FIP
    Joe Smith SD 3.00 3.50
    John Doe Col 4.00 3.50

    I can tell by this that both pitchers were really equally effective, when you strip away their luck, defense, strand rate, and park factors. Is this correct?

    So if I’m a MLB GM considering a trade or free agent signing, this gives me a good measure of the true talent level of these pitchers.

    But if I’m a fantasy baseball player, and I want to know who’s going to have a better ERA for 2013; and I know that both pitchers will still be pitching for SD and COL, respectively, then I would want to know what their FIP would be for someone pitching at SD or COL.

    So where do I find that information? The FIP- apparently is park and league adjusted, but it gives a number that doesn’t equate to an ERA. For example, Kershaws FIP- says “78”. The only instruction I’ve seen is that 100 is average, and the lower the better.

    So, in the above example, if Joe Smith’s FIP- said “78” or “52” or “123”, how do I use that number to adjust his ERA of 3.00 or his FIP of 3.50?

    Comment by Larry — November 23, 2012 @ 12:41 am

  23. Well, BABIP is a significant component of ERA (0.528 correlation in my main sample). FIP, however, doesn’t include BABIP or its inputs as part of its equation at all, as it goes along with Voros McCracken’s idea of BABIP being unpredictable. I think there’s room for a little bit of an improvement to that, as popup-inducing pitchers (e.g. Jered Weaver, Chris Young, Barry Zito, and Matt Cain) can be expected to have good BABIPs more often than not. So, what I’m saying is that in your example, if pitcher A is like Weaver, perhaps he actually is better than his FIP indicates, thanks to his legitimately lower BABIPs (all of those guys I mentioned have FIPs at least 0.4 higher than their ERAs, and their xFIPs are much higher than even their FIPs). But I’ll have to mess with that and see if I can come up with some effective method of integrating this BABIP stuff into ERA projections (as SIERA tries to do, though it’s no better than FIP, it seems to me).

    Regarding FIP-, as you saw, it’s worse at predicting future years for a pitcher than plain FIP. That’s because pitchers usually play for the same team they did in the previous year. I’d say that FIP- is more useful in the context of arguing which pitcher is better than which than it is for making projections.

    So, if they’re playing for the same team as the previous year, just use their straight FIP (better yet, a weighted average of more relevant FIPs from earlier years). But if the player pitched in Colorado one year and San Diego the next, I imagine you could use these:
    … which have the basic park factors for the Rockies at 113 and for the Padres at 92. So, I guess you could just multiply his previous FIP by (92/113), if you’re looking for the quick and dirty method. If you want the [probably] better method, you’d use multi-year averages for park factors, and apply the specific park factor ratios for HR, BB, and SOs directly to the pitcher’s numbers within the aforementioned FIP formula. I haven’t really tested any of this out, so I can’t vouch for the accuracy of those methods, though.

    I’m a bit skeptical, by the way, as to the reliability of park factors (see how much they vary from year-to-year, inexplicably). I think they’re really measuring the teams’ tendencies in addition to the parks themselves. I haven’t looked into it much, but I’d probably want to give a bit less weight to them.

    Comment by Steve Staude. — November 23, 2012 @ 4:34 am

  24. Thanks again, but I’m still confused.

    1) You said “FIP doesn’t include BABIP…” but I thought a big part of FIP, xERA, SIERA and all the other measures were to normalize BABIP to try and eliminate the luck factor?

    2) I thought that FIP shows what a pitcher should have done with a neutral team in a neutral park, and that FIP- shows what they should have done with their specific team in that specific park. But according to what you just wrote, it looks like I’ve got that backwards?

    Comment by Larry — November 23, 2012 @ 12:37 pm

  25. 1) Take a look at the FIP formula from my first reply to you — it ignores hits (other than home runs, which BABIP ignores). It’s based on the assumption that pitchers should have the same BABIPs each year. SIERA and tERA don’t make that assumption — they look at the batted ball profiles, to try to figure out a pitcher’s “true” BABIP, as a component of their “true” ERA (though they apparently don’t do a great job of that). xERA, which you brought up, (but I don’t have the numbers for) apparently looks directly at the pitcher’s hits allowed as a factor, so it’s pretty much the polar opposite of FIP (unless we’re talking about different versions of xERA).

    2) Yeah, you had it backwards; the formula for FIP doesn’t include league averages or park effects, but FIP- does. FIP- is the pitching equivalent of OPS+ or wRC+, as it says here:

    Comment by Steve Staude. — November 23, 2012 @ 3:51 pm

  26. I’m a little late in reading and responding to this article, but both this (and the second installment) are excellent! Thanks for including the “gratuitous” graph at the end, along with the correlations of various pitching stats.

    It’s great to see how many of the individual seasons fall within +/- 10 points of .290, which has the highest frequency. In the correlations, it’s instructive to see just how much stronger YTY correlations on things like GB/FB and contact rates are than BABIP. Again, this is great work!

    Comment by Michael Mitchell — December 19, 2012 @ 11:27 am

  27. Thank you very much, Michael! Stay tuned, next I’ll be talking about future BABIPs and applying all this to ERAs.

    Comment by Steve Staude. — December 20, 2012 @ 10:25 pm

Leave a comment

Line and paragraph breaks automatic, e-mail address never displayed, HTML allowed: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>


Close this window.

0.334 Powered by WordPress