from Amazon
Monday, March 08, 2010
Pre-Introducing Batted Ball FIP
A pitcher’s performance can be split between what includes his fielders and what doesn’t. The latter is captured by the component called FIP (fielding independent pitching). The former includes hits and outs, putting the pitcher at the mercy of his fielders and his parks. We need a separate metric to capture this component of pitching.
A few years ago, I planted the seed of what a batted ball FIP would look like: rather than looking at the result of the at bat in terms of hits and outs, we’d instead look at whether the batted ball was a groundball or flyball.
Many analysts have done many good things with batted ball based component ERA. I’m here to set the bar. Not high, but low. As with Marcel, I’m not interested in creating the best possible system. I’m instead interested in creating the least acceptable system that is relevant. FIP was one such system. It works fantastically well because:
1. it uses limited amount of data (K, BB, HR, IP)
2. combines them in an easy and easy-to-remember fashion
3. produces such good results that it gets us most of the way there
DIPS and BaseRuns-DIPS is a better system than what I have done. And if you can, choose the better system. But, FIP is what is ubiquitous, not DIPS. Everyone else can fight and claw their way from my low point to the top. I’ll be happy to stay one cut below them.
And so, I now introduce Batted Ball Fielding Independent Pitching (bbFIP). I’ll also show you how you too can create a metric as easily as possible. Your first step is to get hold of data. Metrics are created to try to capture existing historical data and try to make sense of it. We’re not here to invent a wheel: we’re here to explain how a wheel turns.
I have in my possession pitcher data from 2002-2009, totalled by pitcher that includes these results:
- strikeouts, walks, hit batters
- groundballs, outfield flyballs, infield flyballs, line drives
- runs allowed, plate appearances
I’ve got 311 pitchers with at least 1500 plate appearances.
The question on the table is: what is the relationship between all those events in the first two lines and runs allowed? Well, we look at the data and try to figure it out. The best thing to do is run a regression, but presuming most of you are like I am, we want to get our hands dirty and try to see the results for ourselves.
Let’s start with the first line. I’m going to exclude intentional walks and include hit batters. Then, I’m going to classify each of my 311 pitchers into 3 groups of walk rates (low, normal, high) and 3 groups of strikeout rates (low, normal, high).
The walk groups are based on these boundaries: .096 walks per PA, .077 walks per PA. If you have less than .077, you are in the low walk group. If you have more than .096, you are in the high walk group. We end up seeing this:
n avgBB avgRA BBclass
107 0.065 4.45 1_Low
102 0.087 4.78 2_Norm
102 0.111 4.76 3_High
We have 107 pitchers in the low walk group, who averaged .065 walks per batter and allowed 4.45 runs per 9 IP. It’s not terribly interesting other than how uninteresting it is. The question is if there is BIAS in the groupings. Is there something that is uncontrolled that is linked to walks to such an extent that the results we are seeing is not just about walks, but about something else? Let’s also include strikeout rates:
n avgBB avgSO avgRA BBclass
107 0.065 0.170 4.45 1_Low
102 0.087 0.168 4.78 2_Norm
102 0.111 0.189 4.76 3_High
Ah-ha, so the high walk pitchers also have alot of high strikeout pitchers. So, perhaps the reason that their runs allowed did not skyrocket high is because they had alot of good strikeout pitchers. Our next step is clear: break down by strikeout groups instead.
We’ll use boundaries of .152 and .192 strikeouts per PA. This is what we get:
n avgBB avgSO avgRA SOclass
103 0.081 0.131 5.12 1_Low
102 0.091 0.169 4.89 2_Norm
106 0.089 0.225 4.00 3_High
Now we see that high strikeout pitchers give up very few runs. This is a huge indicator. Walks and strikeouts are always used together, so let’s include BOTH groups: those that break down by walks and by strikeouts. We get this:
n avgBB avgSO avgRA SOclass BBclass
34 0.064 0.223 3.70 3_High 1_Low
29 0.087 0.221 4.03 3_High 2_Norm
43 0.111 0.230 4.22 3_High 3_High
30 0.066 0.167 4.60 2_Norm 1_Low
32 0.087 0.169 4.84 2_Norm 2_Norm
40 0.112 0.170 5.14 2_Norm 3_High
43 0.065 0.129 4.95 1_Low 1_Low
41 0.086 0.130 5.28 1_Low 2_Norm
19 0.108 0.137 5.16 1_Low 3_High
The best group is, obviously, the pitchers who strikeout the most and walk the least. Those 34 pitchers averaged .064 walks, .223 strikeouts, and allowed 3.70 runs per game. The worst pitchers either gave up alot of walks or struck out few batters.
Before we go off an run a regression of which most of you might not even believe, let’s look at the numbers. Specifically, let’s look at the first three rows: they are all from high strikeout pitchers, and their strikeout rates are around .22 to .23 per PA. The range in runs allowed goes from 3.70 to 4.22, and that’s based on the range of walks allowed of .064 to .111. So, we can roughly say that, for this group of high-K pitchers, giving up an extra .047 walks per PA (.111 minus .064 equals .047) leads to an extra 0.52 runs per 9 IP (4.22 minus 3.70). Or, more generally speaking each .1 walk per PA adds 1.10 runs per 9 IP (.52/.047*.1).
Let’s look at the second group of pitchers, those wth a K-rate of around .170. For those pitchers, the walk rates go from .066 to .112 (difference of .046 walks) and that leads to a change in runs from 4.60 to 5.14 (difference of .54 runs). Well, these numbers are fantastically similar to the previous group, so we end up with a similar result: each .1 walk per PA adds 1.17 rus per 9 IP.
Finally, in the third group, .043 walks per PA adds, well, it’s not as clear because we have a little blip in there. We’ll get back to that in a second.
We can repeat this with the walk groups. If you look at the first row of each group, you will see they are all the low-walk groups, and they all have around the same number of walks (.064 to .066). The differentiator is the strikeouts (.223 down to .129) and the runs allowed moves the other way (3.70 to 4.95). A difference of nearly .100 strikeouts leads to 1.25 runs per 9 IP (or more precisely 1.32 runs per 9IP for each 0.1 K per PA).
If we look at each of the second rows, we see the BB rates are stable, whike the K rates go from .221 to .130, for a change in runs allowed of 4.03 to 5.28, or a rate of 1.37 runs per 9IP for each 0.1 K per PA.
Finally, the last row will give us 1.00 runs per 9IP for each 0.1 K per PA change.
Overall, we can see that, generally speaking, each 0.1 K per PA moved the runs allowed by around 1.0 to 1.3 runs per 9IP and each 0.1 BB per PA moved the runs allowed by around the same amount but in the opposite direction.
And this is where we run our first regression. Rather than doing everything that I just did, in terms of binning and finding differences, etc, we run a regression. This is what the regression was designed for. When we run a regression of the two independent variables (strikeouts and walks per PA) against the dependent variable (runs per 9 IP), we end up with this regression equation:
RA = 5.84 + 11.8*BB - 12.6*SO
As you can see, this is exactly what we should have expected: each 0.1 strikeout should have added around 1.3 runs (or each 1 strikeout would add 13 runs).
As interesting as the one walk impacts one strikeouts finding is, the more interesting finding is that the correlation coefficient is an r=0.75. This is a very high number, as you will soon see.
Now, as I said, I’m a simple guy. I don’t want to see “11.8” and “12.6” in there. They are close enough that perhaps we should just look at BB-SO (treating them both equally) and letting the regression figure out the proper weight. And this is what we get:
RA = 5.76 + 12.5 * (BB - SO)
(If you need it in the form of ERA, multiply the resultant answer by 0.92, because 92% of runs allowed are earned.) And there you have it, the first step in figuring out bbFIP. And, incredibly, we could actually stop right here, and consider NOTHING ELSE about a pitcher, and we’d be really close to capturing his performance. Really, you ask?
Here is a chart that plots a pitchers Strikeouts minus Walks on the X-axis and runs allowed per 9IP on the y-axis.
That’s what an r=.75 chart looks like. And this was done by looking only a pitcher’s walks and strikeout and nothing else. I know, pretty cool. Anything we do beyond this point is going to be gravy. I’ll get to part 2 next time.
This is *great*.
I take it runs allowed was chosen because you are explicitly looking at the component of pitcher performance that includes fielding?