I have always loved pitchers’ duels. One of my favorite childhood baseball memories is watching Curt Schilling throw a complete game shutout for the Phillies in a 20 win against the Blue Jays in Game Five of the 1993 World Series, with the Phillies facing elimination. I was only 12 years old at the time, and I did not know anything about sabermetrics, but Schilling appeared majestic as he pitched yet another brilliant start in what would become a magnificent playoff career. He only surrendered five singles that night and extended the series one more day.
When I finally did learn about Defense Independent Pitching Statistics (DIPS) and Voros McCracken’s discovery that pitchers have little control over batting average on balls in play, I found it mildly disappointing to look back at that box score and discover Schilling actually had rather mediocre DIPS numbers that night—just six strikeouts and three walks. However, I still felt confident in my memory that I had seen a great game as a youngster, and I convinced myself that the lack of home runs that Schilling allowed were evidence that he was unhittable that night. Unfortunately, I later discovered that pitchers have little control over the rate of home runs allowed per fly ball, and that Schilling actually generated just 10 ground balls that night on 24 balls in play. Though DIPS Theory finds Schilling’s entire career amazing, his numbers from that alleged gem appeared pedestrian even on that front.
Some fans probably dislike DIPS Theory because it takes the glamour away from memories like these, but I find myself appreciating individual pitching performances more now that I have a better understanding of the game. Seeing a pitcher like Tim Lincecum strike out 14 hitters in his playoff debut adds to the majesty of the 2010 postseason, rather than takes away from Roy Halladay’s playoff nohitter the day before. The knowledge of what good pitchers do and do not control has given me an ability to watch a brilliant pitching performance critically and appreciate the masterful performances more.
Elementary DIPS theory is well established. Pitchers have a lot of control over strikeout, walk, and groundball rates, as evidenced by the yeartoyear correlation of this statistic, but they have poor control over batting average on balls in play (BABIP). While we have known for some time that pitchers also have little control over HR/FB, statistics like FIP remove all of the variance from BABIP and reproduce an ERA with a leagueaverage BABIP for all pitchers, but leave in all of the variance in HR/FB. This treats HR/FB as a skill and hides the fact that home it is a statistic that pitchers show even less ability to maintain than their BABIP. Looking at home runs per outfield fly ball is important, because some pitchers do have some tendency to allow more popups than others, even controlling for team.
Below I list the yeartoyear correlation of different statistics for 549 pitchers with at least 300 balls in play in consecutive seasons from 200310. For BABIP and HR/FB, I give numbers net of team performance (because defense and parks affect these numbers similarly for pitchers on the same team).
Statistic 
Yeartoyear Correlation 
Strikeouts per Batter Faced 
.797 
Walks per Batter Faced 
.667 
Ground balls per Ball in Play 
.839 
Batting Average on Balls in Play (net of team) 
.122 
HR/FB (including popups; net of team) 
.113 
HR/FB (excluding popups; net of team) 
.075 
Pitchers actually appear to have more control over the rate of hits allowed on balls in play than they have control over the rate that outfield fly balls go for home runs. Yet statistics like FIP remove the luck from BABIP, but not HR/FB, because it is known that the rate of HR/FB is a little persistent. However, since BABIP is even more persistent, this does not make much sense. Statistics like SIERA and xFIP that neutralize the number of HR/FB do a better job of explaining reality when you have only a limited amount of information about a pitcher. Tom Tango has often argued that several seasons of data will show patterns in HR/FB, and that FIP will outlast xFIP when predicting ERA for pitchers with several years of data available. The problem is that excluding BABIP for several years of data by using FIP will ignore the control that pitchers have over BABIP as it reveals HR/FB skill, and ERA itself may be better at showing a pitcher’s skill level with several years of data than FIP will.
Somewhat interesting was that popup rate has a high yeartoyear correlation itself (.582). But what was surprising was that even after adjusting for team (and therefore park and foul ground area), the rate of popups per fly ball also had a pretty high yeartoyear correlation (.328). This suggests that inducing popups is a skill beyond being a product of being a flyball pitcher. Therefore, SIERA’s inclusion of outfield fly balls and infield popups together may be something worth revisiting as we learn more about pitching.
All numbers in the below table are computed relative to total team numbers, and are for all pitchers with at least 100 balls in play in consecutive seasons between 2003 and 2010 (a sample size of 1459).
Correlation 
Popups per ball in play(same year) 
Popups per ball in play (next year) 
Fly balls per ball in play (same year) 
Fly balls per ball in play (next year) 
Popups/(Popups + Outfield Fly balls) (same year) 
Popups/(Popups + Outfield Fly balls) (next year) 
Popups per ball in play(same year) 
1.00 





Popups per ball in play (next year) 
.582 
1.00 




Fly balls per ball in play (same year) 
.542 
.494 
1.00 



Fly balls per ball in play (next year) 
.571 
.494 
.635 
1.00 


Popups/(Popups + Outfield Fly balls) (same year) 
.838 
.356 
.044 
.293 
1.00 

Popups/(Popups + Outfield Fly balls) (next year) 
.317 
.058 
.085 
.072 
.328 
1.00 
However, the rate of popups per ball in play does not have a significant correlation with the rate of home runs per outfield fly ball (.043). The importance of making an adjustment to SIERA seems minimally important, as the correlation between nextyear’s ERA and this year’s popup rate is low as well (.028). On the other hand, regressing next year’s ERA on this year’s popup rate and this year’s outfield flyball rate shows an interesting effect.
Next Year’s ERA = 4.96 – 9.77*(net PU%) + 7.71*(net FB%)
The net popup rate coefficient is significant, with p=.022. This suggests that differentiating popups and fly balls may be more useful and could be used to enhance SIERA.
Although the rate of popups per ball in play does not correlate with the rate of HR/FB, this does not mean that we cannot learn something important about this statistic. In fact, part of the strength of SIERA comes from the fact that it implicitly models the rate of HR/FB similarly to how it implicitly models BABIP.
I checked the correlation of HR/OFB with DIPS statistics:
Statistic 
Correlation with HR/FB (100 balls in play) 
Correlation with HR/FB (300 balls in play) 
SO/PA 
.116 
.083 
BB/PA 
.001 
.018 
GB/BIP 
.028 
.003 
The reason that the strikeout coefficient in SIERA is so large is that not only do strikeouts lead to lower ERAs in and of themselves, but they also correlate with lower BABIPs and lower HR/OFB rates, which likewise correlate with lower ERAs.
Running a regression (on all pitchers with 100 balls in play or more) of HR/OFB rate on all three statistics does not show a statistically significant coefficient for anything but strikeout rate:
Variable 
Coefficient 
PStatistic 
SO/PA 
.100 
.000 
BB/PA 
.014 
.653 
GB/BIP 
.003 
.747 
Constant 
.013 
.037 
The equation of the regression with only strikeout rate is:
Net home runs per outfield fly ball = .016 – .100*(SO/PA)
In the above regression, the pstatistic on strikeout rate is also less than .001.
yesterday’s article both show the benefit of using regression analysis to study ERA. Pitchers clearly exhibit far more control over defenseindependent pitching statistics, but they still do have some control over BABIP and HR/FB. While looking at statistics like FIP removes the noise inherent in these two metrics, they also remove the skill. Since pitchers’ BABIP and HR/FB skills are significantly correlated with their DIPS skills, running a regression actually controls for these effects and gives an accurate reading of pitchers’ true skill levels with a unique methodology that picks up on factors that other measures do not.
Overall, this article andThank you for reading
This is a free article. If you enjoyed it, consider subscribing to Baseball Prospectus. Subscriptions support ongoing public baseball research and analysis in an increasingly proprietary environment.
Subscribe now
With respect to BABIP and groundballs, I found Jamie Moyer's season last year very interesting. There was certainly a difference between those games when he got good results vs. bad. I bring him up because I watched a lot of his games last year and have been trying to figure how to square what I *thought* I saw with what the advanced metrics told us about his value to the Phillies.
We talk about a very low BABIP as being unsustainable, and rightly so, I think, but it seems to me that we make a mistake if we look at a low BABIP and simply conclude "he was wasn't *really* pitching that well". In those games in which Moyer posted the best results last season, he did indeed (somehow) induce a lot of weak contact, weak grounders and flies that are easy to field. Obviously weakly hit balls can find holes and be misplayed too, so luck remains a factor. Still, it seems to me that DIPS theories don't tell us much about any one game, but rather about the persistence of certain tendencies. With a pitcher like Moyer, he needs to have very good command to be successful (and to "get" the outside strike), but when he does have that command he *really* did pitch that wellweak contact mattersbut if he's slightly off, he gets hit much harder.
Am I making any sense? If so, is there anything to what I'm suggesting?
FIP takes no position on the amount of skill level captured in any metric. The single and sole purpose of FIP is to capture the HR, BB, HB, SO observed results (a subset of a pitcher's results) and expresses it as a single number. (And, for ease of use, scale it to ERA.) This is no different than OBP treating BB and HR identically.
That FIP includes persistent results like SO and lessthanpersistent results like HR is irrelevant. That's exactly the case with OBP as well, as it includes a persistent result like BB and lessthanpersistent results like singles.
If you want to know how the observed results, things that actually happened, is associated to runs, then FIP gives you that. SIERA will not give you that. That's not a knock on SIERA, but neither are we going to hold it to the standard that it doesn't tell us what actually happened, when it doesn't purport to do that to begin with.
Observed results is a combination of the true talent level, other biases, and random variation. If you want to know the pitcher's true talent level inside that observed result, then clearly you have to do something to FIP in order to use it for that.
Back to Matt's article now...
If you want to know the pitcher's true talent level inside that observed result, then clearly you have to do something to FIP in order to use it for that.
Matt isn't trying to find a pitcher's true FIP talent. He's trying to find a pitcher's true ERA talent, and these articles are about the adjustments made to SIERA that may allow it to be used for that.
It's possible that there are pitchers who adjust their style to fit a ballpark. But that sounds suspiciously similar to the old claim, "a good pitcher throws to the score" (cough cough jack morris). In that case, the onus would be on the person making the assertion to prove it. I just doubt that there's much of an effect from it.
"Pitchers clearly exhibit far more control over defenseindependent pitching statistics, but they still do have some control over BABIP and HR/FB."
I didn't see the evidence for the latter part of that sentence, nor any data on how much control the most effective pitchers might consistently show over BABIP. Thank you.
If r=.10 when BIP=400, then r=.50 when BIP=3600.
Unless you have a systematic bias, you can get r to approach 1 on almost anything, as the number of trials approaches infinity.
The sample size issue that Tango is talking about with HR/FB is cancelled out somewhat by the fact that HR/FB is lower than BABIP. HR/ofFB is around .130 on average. BABIP is around .300. About 28% of BIP are oFFB. So if you take a pitcher with N BIPs and .28*N ofFBs, then your standard deviation due to randomness is:
For BABIP: sqrt{(.300)*(1.300)/N} = .46/sqrt(N)
For HR/ofFB: sqrt{(.130)*(1.130)/(.28N)} = .64/sqrt(N)
So there is a slightly larger standard deviation due to randomness, about 1.39 times as large for HR/oFFB. But the correlation for BABIP is .122 vs. .075 for HR/ofFB, which is 63% larger. So we're still looking at BABIP having higher variance in skill level (***assuming I did all that correctly***)