THE BOOK--Playing The Percentages In Baseball

I dont doubt that at all. It’s a great and simple formula. Value comes from simplicity

Alas, i’d like to see where my modified version of FIP ends up. Its the same formula as xFIP, but it uses expected home runs and normalized line drive and popup rates

Interesting that WPA “Clutch” has a 0 Y2Y correlation.  I have been thinking about the ramification of guys like Ryan and Palmer with +/- 6 or 7 career Clutch scores.  Now I am not so sure if it is worth looking at at all.  Maybe we can see some correlation over many seasons.

It looks like the color codes missed quite a few entries with an r of .40 or more (+ or - of course).

1/Jeffrey:
That sounds a bit like tERA, right?

Would be nice to see how (K-BB)/PA does.

Guy: no question it would be between K/PA and FIP.  Given that BB/9 by itself had almost no correlation, it likely won’t move the needle at all.

This is what happens when you see the threshold at 162 IP: you are not going to have a guy with poor walk rate AND poor K rate.  The K rate really ends up being the driver.  It’s a bias issue if you get BB rate having no correlation to runs allowed.

Tango/6:
That’s not quite true. There is a +.14 correlation between BB%(Year 1) & K%(Year 2). That’s probably what’s driving the ~0 correlation between BB% and K%. I’m guessing that if you include relievers, you’d even get a mildly stronger positive correlation between BB%(Year 1) & ERA(Year 2), because the guys who have high BB% & high K% who stay in the league are often relievers. I’d still guess (K-BB)/PA does better. The reason K/PA drives everything so much is the same reason its average derivative is so much highest in SIERA than FIP-- K% is correlated with low BABIPs and low HR/FBs.

In this year’s Hardball Times Annual, my fielding article includes the shortest ‘linear weights’ formula for pitchers: (SO - 2*BB)/3.

This provides a good estimate of runs ‘saved’ versus average.  It’s particularly useful for relievers and those with small sample sizes.

The multiple for BB--2--could be modified to equal the precise ratio of SO to BB, which has recently drifted above 2.

The implicit run-weight per SO and BB is the same.  Lindsey-Palmer Markov models indicate about .33 runs allowed per BB, but only .25 to .30 runs saved per SO.

However, the DRA regression coefficient for recent years for SO is about .33 as well, and can be explained by the fact that each SO is associated slightly with run prevention beyond the immediate out (that is, fewer HR/contact and fewer H/bip).

MAH:  If BB and K have about the same absolute value (agreed), then why do you use the K-2*BB metric?

It looks like he does it to “force” the thing at zero.

In any case, the run value of the BB and K are about .02 or .03 runs apart or so.  Regression also shows that the best fit of BB/PA and K/PA to RA9 to be the same coefficient for BB/PA and K/PA.

Hence, (K-BB)/PA is what works the best, both theoretically, and in reality.

Yeah, the SIERA coefficients are very similar to (K-BB)/PA for the average pitcher. If you take the derivative of SIERA with respect to each peripheral, and plug in the league average for each peripheral, and then multiply by league average PA/IP, you actually get FIP-ified coefficients of about 2.9 for both SO & BB, and a much less steep coefficient for FB than xFIP would imply (about 0.7 for SIERA vs. 1.3 for xFIP), so you really do get very close by doing (K-BB)/PA.

I would guess that (K-BB-.09)/PA would do better than (K-2*BB)/PA but I’m curious what MAH has to say on his reasoning before I assume that.

It’s SO minus 2*BB because you’re measuring runs saved relative to average.  If SO ~ BB, but the the league SO/BB ratio is 2:1, then you won’t be ‘ahead’ as a pitcher unless your SO exceed 2*BB.

@Tango good call out on the selection issue. This was partly on purpoae as i wanted to see what the relationship was for pure starters (i.e.~ >= 27 starts per year), but you are right that it will drive down th correlationof walks.

@dave Pretty sure the .40 that were not shaded blue was because they were rounded up from .39 and I set the formatting floor at .40.

#12, using linear weights presupposes the ratio doesn’t matter, only the difference.  If adding a strikeout and subtracting a walk are of equal value, then going from 6 K/9 and 3 BB/9 to 7 K/9 and 3 BB/9 is worth just as much as going from 6 and 3 to 6 and 2, even though the ratio is worse.  Likewise, adding 3 strikeouts and 2 walks would be a net positive, so 9 and 5 would be better than 6 and 3 even though the ratio drops below 2.

Setting the coefficient for walks at twice that of strikeouts implies linear weights values where the walk has twice the impact of the strikeout.  No matter what the ratio is, a loss of 2 walks would offset the gain of 1 strikeout, so they can’t have the same linear weights value in that formula.

You would be better off centering the value at zero by adding in a constant term like Matt/#11 suggests.

Bill#13, no there are values over .40. I think even maybe a couple over .50 that aren’t shaded

Kincaid is correct, as usual.  (Always?) The stable term is K-BB per PA.

If

[1] You want a linear formula for pitchers estimating runs saved or allowed relative to league average based solely on their SO and BB,

[2] The league average ratio of SO to BB is 2:1,

[3] the regression weight in a comprehensive DRA model of team defense in the 1990s and 2000s in runs per BB is about .33 (as per Lindsey-Palmer Markov models) and also about .33 per SO (because SO associated with fewer HR and fewer non-HR hits after taking into account AB minus SO), then

[4] you arrive at the (SO - 2*BB)/3 formula.  Let me know if you would like me to walk through the arithmetic.

If you don’t want a linear formula, or you want some kind of absolute formula not centered around zero, then something else will be necessary.

But in terms of providing the absolutely simplest formula for pitcher linear weights runs, I think this is the best, and hard to believe, never before proposed.

MAH, what’s wrong with (SO-BB-.09)/PA? It’s linear, just with a constant. The reason that I prefer this to (SO-2*BB)/PA is that a pitcher with 20% Ks & 11% BBs is going to generally allow fewer runs than one with 10% Ks and 4% BBs, but your formula suggests otherwise.

---” Let me know if you would like me to walk through the arithmetic.

*******************

Ok, I’ll bite.

Michael: it has been proposed before, and it was wrong then, and it’s wrong now.

In your case, you are saying that these two pitchers are equivalent:
26% K, 13% BB
6% K, 3% BB

The reality is that the 26% K, 13% BB pitcher is equivalent to 20% K, 7% BB pitcher.

The league average RATIO is irrelevant.  The only thing that counts is the league average DIFFERENCE.

Whatever math you present, if it confirms your opinion, means that either you did the math wrong, or you are going through mathematical gyrations.

Have at it, and we’ll see where you went wrong.

Go back to what I wrote.

I wanted a linear formula “based solely on SO and BB.”

The other formulas are based on SO, BB, _and BF_ and/or the other analyses are based on percentages (SO% and BB%), which rely upon BF data.

If you’re watching the game on TV and you see a pitcher’s SO, BB, and BF, and you want to calculate his linear weights, you absolutely do get

.33*(SO - BB - .085*BF). 

This can be derived, using 2007 MLB data when SO:BB was almost exactly 2:1 (and IBB ~ HBP) and DRA pitcher formula

.33*SO|bf - .33*SO|bf - 1.50*HR|bf-so-bb
+ .80*A1|bip + .80*IFO|bip - .40*WP|rob.*

Just the first two terms are equal to

+ .33*(SO - BF*(lgSO/lgBF)) and

- .33*(BB - BF*(lg/BB/lgBF)),

which equals

+ .33*(SO - BF*(.170)) - .33*(BB - BF*(.085)),

or

.33*(SO - BB - BF*((.170) + (.085))),

which is

.33*(SO - BB - .085*BF).

But baseball telecasts rarely display BF.  And even if they did, I would find it a pain to multiply anything in my head by .085.

If you _don’t_ know anything but the pitcher’s SO and BB, the ‘arithmetic’ is simply that you ‘need’ MORE than twice as many SO as BB to be ‘ahead’ of expectations, or the SO need to be more than 2*BB.  Since each SO (and BB) is worth .33 runs you just take (SO - 2*BB)/3.

It would be an interesting exercise in Excel to show the difference between the results of

.33*(SO - BB - .085*BF) and .33*(SO - 2*BB)

for 2007 pitchers. Undoubtedly for outliers with low or high levels of allowing batters to make contact would diverge.

*The run weight is too high because the ‘denominator’ of opportunities should be WP (including PB) given total BF with one or more runners on base.  This should bring the run-weight down to .25-.30.

Michael: ok, if you give yourself that constraint, then that’s fine.

Search my blog for lego-metrics.  It does close to what you just did.

From MAH #17:

But in terms of providing the absolutely simplest formula for pitcher linear weights runs, I think this is the best, and hard to believe, never before proposed.
*****************

To me, your formula only works by accident, when the lg K/BB ratio happens to be 2.0. But there was a time when it was 1.0, and then 1.5, and then 2.0, and now it’s 2.2 or so. Who knows where it will go in the future?

Perhaps using the ratio is the best solution if one does not have PA or IP for a pitcher (with Dave’s caveat that it only works in the current 2:1 environment).  But when would that ever be the case?  I see the case for parsimony—e.g. using K-BB if it’s 98% as good as FIP—but why exclude PA/IP if that allows you to be considerably more accurate?

As I noted in the lego-metrics thread, that what Micheal is doing is creating a fun toy:

http://www.insidethebook.com/ee/index.php/site/comments/lego/

It should not be used for anything seriously analytical.  The idea that you ONLY have SO and BB, and you don’t have IP or PA can only happen in something like high school or something like that.  And even then, you’d also need to know the league SO and BB numbers. 

We know the run values of the BB and SO.  We know that if you need a shortcut, that you need a third number (IP or PA).  And that provides a fantastic base.

Limiting yourself to BB, SO, and the league ratio is, well, limiting.

The whole point of (SO - 2*BB)/3 is that you can calculate it in your head while watching a major league pitcher warm up on TV or while skimming a hardcopy or online encyclopedia for a simple DIPS-esque run estimate for pitchers.  No need to limit it to your high school games (for which it would almost certainly be invalid anyway). 

In that respect it is clearly better than anything in the lego thread.  And it’s unique.  And never proposed before.  Sure, you can add BF.  But at that point, you might as well use the computer, in which case using only SO, BB, and BF would be, well, limiting.

A few more points.  The reason “we know the run values for BB and SO” for this purpose is because of Defensive Regression Analysis.  If we relied on the Lindsey-Palmer Markovian model weights of .27 for SO and .33 for BB, that is, if they weren’t both associated with +/-.33 runs saved, it impossible to create a ‘calculuate in your head’ formula.

All that happened when I ran DRA for THT 2012 article is that I saw the weights were roughly equal and rougly one-third during the sample seasons used (1989-99; 2002-2010).  The ‘extra’ weight for SO could be accounted for by the fact that it was to a non-trivial degree associated with fewer HR given BF - SO - BB and also somewhat better BABIP.  This statistical relationship, coupled with the fact that there happened to be roughly twice as many SO as BB during that time frame, caused the formula to pop out.

Smyth, yes, it only works “by accident” during the past decade or two when the SO:BB ratio has been approximately 2:1.  I believe I specficially mentioned in The Hardball Times Annual 2012 that the SO:BB ratio has been creeping up in the last couple of years, making the formula less accurate.

Wizardry’s appendix on pitchers specifically addresses the shifts over time in the SO:BB ratio, and emphasisized how it was consistently about 1:1 until the 1950s, when it jumped to 2:1, fell a bit in the 1970s and ‘80s.  For pitchers from the 1920s through the early 1950s, our little lego formula is even simpler:

(SO - BB)/2.

Yes, that’s because both SO and BB were both much more strongly associated with HR avoidance/allowance and H-HR avoidance/allowance, resulting in runweights of +.5 and -.5 respectively.  So again, when you’re flipping through an old baseball encyclopedia, you can quickly size up the pitchers.

I don’t know why you say it’s “never” been proposed before.  I mean, it’s the same stat in anything.  If league average OBP is .333, then guess what also works:

times on base - outs/2

With SB and CS, you do SB - 2*CS, because that’s around the league average.

If the league ratio is 2 K per 1 BB, then naturally you get:

K - 2*BB

I’ve seen that K - 2*BB countless times.

I’ve seen people do W-L (to do above average) and I’ve seen people do 2*W-L (to do above replacement).

Anytime the baseline rate is .333 or .667, you always do 2*X - Y or X - 2*Y.

It’s just not that big a deal.

***

Now, if you are flipping through the encyclopedia, or looking at the screen stats while the pitcher is warming up, now you are doing two things:

1. Intentionally ignoring IP

2. Subliminally using IP as a threshold, to make sure that you can do SO - 2*BB to give meaningful results.

For example, a pitcher has 30 K and 10 BB.  You are going to react differently to that if it was done in 25 innings or if it was done in 50 innings.  You are not just going to say “+10”.

So, I don’t see the benefit here.

***

Now, is SO - 2*BB more stable than SO/BB?  Well, that depends on how many innings you have.

Somewhere at around 150-200 batters faced is where you’d draw the line.  So, if you have less than 40 innings, using SO - 2*BB makes more sense.  With more than 40 innings, then the ratio is more telling.

Except you’ve intentionally excluded the use of IP and PA.

40 innings is also about 40 SO+BB.  So, you can use that instead. 

In the above example, if you have 30 K and 10 BB, you can go either way, using the ratio or the differential (or actually, using both).

***

But, in no way would I use this as anything other than a toy.

All these are going to give you +10 (K first, BB second):

10,0
30,10
50,20
70,30
90,40
110,50
130,60
150,70
170,80

Now, you really have to know how many innings, or how many games a team played.  Because a 50/20 might be more impressive than 170/80.  Or it could be less impressive.  I can’t tell.

If you do K-BB, which you should, you have +30 for 50/20 and you have +90 for 170/80.  If the second pitcher threw 3 times more innings than the first, then they’re equivalent.

Any use of K and BB, without the use of IP or PA can only be considered a toy.

Tom, you really are Petty.

Michael: there is NO reason to take this personally.  And even less reason to characterize me personally as well.

You cross the line when you make it personal.

"Michael: it has been proposed before, and it was wrong then, and it’s wrong now. * * * Whatever math you present, if it confirms your opinion, means that either you did the math wrong, or you are going through mathematical gyrations. * * *
Have at it, and we’ll see where you went wrong.”

But you never say who previously proposed “(SO - 2*BB)/3”; you just state that it’s been proposed before.  Rather than inquire about the formula, you dismiss it out of hand as “wrong” and then add, “We’ll see where you went wrong.” NOT “let’s figure this out,” but “we’ll show you you’re wrong.”

Then once I explained the premise and defended the formula, you shifted tack.  Since it was no longer “wrong,” it had to be trivialized by relegating it literally to “high school.”

And you still felt it necessary to devote two more blog posts to show (1) it ain’t new, and (2) it’s only a “toy.”

Look, I only bothered writing because it is a surprisingly simple and nifty formula.  I never meant it to be anything other than a quick thing people can calculate in their heads.  I also think it can be helpful to people who are brand new to statistical analysis in baseball as a less forbidding stat that cuts through the clutter when you see a reliever take the mound and his ERA pops up.

That you have devoted so much energy to ‘proving’ to your audience that this little thing (1) ain’t new, (2) is “wrong”, and (3) without any use you’re willing to acknowledge, suggests to me that you were the first in this exchange to take things personally.

Michael:

Don’t infer something.  And don’t then use that inference to make a personal attack.

I did not take it personally, and don’t tell me that I did take it personally.  You can’t tell me how I feel.

You crossed the personal line.  That’s the only thing that’s a matter of fact. 

The decent thing to do is to either apologize, or, just move on.  Incredibly, you are now justifying your words.

I have zero desire to talk about personal matters.  None whatsoever.  Stop justifying your actions.

Now, if you still have the need to talk about personal issues, email me.  It’s pollution to talk about your personal issues about me on my blog.

The formula is “wrong” in the sense that it forces you to give walks twice the impact they actually have.  The two key aspects of linear weights are that they center around zero, and that each event is weighted based on its average run impact.  Trying to approximate linear weights using only total strikeouts and walks lets you center the weights around zero, but in doing so, the weights no longer represent the impact of each event.  They are no longer telling you the impact of each event, but rather the weight you have to give it to center at zero.  If that weight ends up being twice what it should be in terms of the run impact of that event, that’s wrong from a linear weights perspective.  It doesn’t mean you are wrong for doing so if you force yourself to limit your inputs to Ks and BBs, but the inputs are severely limited from calculating linear weights.

It’s neat that you can use a simple formula that is easy to calculate in your head by limiting the inputs to just Ks and BBs, but Tango is right that throwing out IP is a huge limitation.  I don’t see the analytic value when just including IP adds so much to the model by letting you set the weights in line with their actual impact. 

I don’t see it being helpful to people brand new to analysis, because you really have to be aware of the limitations when you use the formula.  Otherwise, you see walks weighted twice as much as strikeouts, and you think walks have twice the impact as strikeouts.  You see the pitcher warming up and figure his value using the formula, and then when he proceeds to walk one and strike out two in the inning, you think he treaded water.  You have to be able to understand that the weights come from the limitations of the inputs, and that their actual values are not tied to those weights.

Having a formula that is easy to calculate in your head can be useful, as long as you understand the limitations.  It can be a quick first step or let you make a rough estimation of how well a pitcher has performed.  I would still want to check Ks and BBs against IP any time I used it to make sure the pitcher doesn’t have an extreme rate, though.

I’ll let the readers draw their own inferences and step aside.

Kin/33: ditto.

MAH: you don’t have to “step aside”.  Just stop giving your opinion about *me*.  Feel free to talk about the issue at hand as much or as little as you want.

I had an inkling that MAH might be getting irritated when he called me Smyth back in post 26, instead of Dave Smyth or Dave.

BTW, MAH, I own and enjoyed your book.

I agree with Dave that Michael’s book is very good and enjoyable.

dave (smyth), thanks for buying Wizardry.  It would be great if you could write an Amazon review.

Tom, if you feel that Wizardry is very good and enjoyable, I’m sure your readers would appreciate a follow-up to this link:
http://www.insidethebook.com/ee/index.php/site/comments/wizardry/#comments

As far as I know, Colin never did write a review for Baseball Prospectus, and as you may know, Bill James wrote on his blog, in response to an “Ask Bill” question inquiring what he “thought of the book,” simply replied that he “would not” read Wizardry.

* * *

I went back to run some real numbers to add some insight to the thread.

First, I looked at the SO/BB ratios from 1893 through 2011, and it appears that there were a few discrete periods in which SO/BB ratios remained fairly consistent from year to year, before a clear drop or increase, often attributable to rule changes:

Years_______avg.___s.d.
2001-11_____2.0____0.1
1986-00_____1.8____0.1
1969-85_____1.6____0.1
1963-68_____2.0____0.1
1955-62_____1.5____0.1
1918-54_____1.0____0.1
1901-17_____1.4____0.1
1893-00_____0.7____0.1

So the time frame when (SO-2BB)/3 would potentially work would be 2001-11. During that time frame there were .52 runs per half-inning and .48 earned runs per half-inning.

I copied pitcher data from the B-R.com site, identified all (560) pairs of succeeding seasons in which a pitcher pitched 162 or more innings in both seasons.  I calculated (1) total “runs saved” relative to the .52 runs-per-inning rate (“RS”), (2) FIP, (3) FIP runs saved (scale FIP to total runs (.52/.48), apply to inning pitched to estimate FIP runs, subtract that from .52*ip) (“FIP RS”), (4) DRA pitcher runs using .33/.33/1.5 for the SO/BB/HR coefficients (“DRA”), (5) gross DRA runs per 9ip and 38 batters faced (“bf”), (6) (SO – BB - .085*bf)/3, and (7) (SO – 2*BB)/3.

Year-1 statistic_______Corr.w/Yr 2________Corr.w/Yr2 ERA
DRA/9ip____________.60_______________.45
DRA/38bf___________.62_______________.46
FIP (per 9ip)_________.59_______________.45
SO/bf_______________.80______________(.45)
(SO-BB-.085bf)/3_____.75______________(.42)
(SO-2BB)/3__________.71______________(.35)

Clearly, among the subset of pitchers who pitch more than 162 innings in consecutive seasons, the best rate stat is SO/bf, because its self-correlation is highest and its correlation with ERA is practically as good as FIP and DRA converted into a rate stat. The “(SO – BB – .085bf)/3” stat yields a net ‘gross’ number of runs saved, so it is not a pure rate stat.  That puts it at a disadvantage with rate stats, because the variation in batters faced from year to year creates more noise.  Yet in spite of this, it predicts next-year ERA just barely less well.

Clearly, (SO-2BB)/3 is a cut below the others in predicting next year ERA. But that’s because it is derived solely from outcomes data, not opportunities data (in this respect (SO-BB-.085bf)/3 is a hybrid, in that it produces a net runs saved estimate, but the estimate is ‘informed’ with BF data).

But here’s the thing, my friends.  What is the correlation between ERA and next-year ERA?  Point-three-six.  Virtually identical with (SO-2BB)/3.  So this “toy” is as predictive of next-year ERA as current year ERA for full-time pitchers.

But wait, there’s more . . . the “toy” gives bottom line net value, not ‘just’ another rate stat, of which there are already many to choose from.  And that net value ‘scales’ with actual runs saved and DIPS-like runs saved:

Year1 stat__________Corr. w/Yr1 ERA______s.d. runs
RS________________1.00_________________18
DRA_______________0.73_________________16
FIP RS _____________0.76________________16
(SO-BB-.085bf)/3_____0.61________________14
(SO-2BB)/3__________0.58 ________________16

Now to address the objections raised in the thread. 

* * *

“It’s just not that big a deal,” because it uses arithmetic seen before. 

I haven’t claimed to invent arithmetic or an entirely new form of baseball formula.  In fact, when I was trying to think of an argument to address the perceived ‘overweighting’ of BB, I considered pointing out that others have used 2*W-L to measure replacement value for starting pitchers.  (Incidentally, (SO-2BB)/3 is set at the .500 replacement level for relievers, for which ERA is least reliable.  Yet another point in its favor.)

What is new is simply the identification of the fact that for recent seasons the relative values of SO and BB, and the relative number of SO and BB, have been aligned in such a way that a very simple calculation can provide a measure of pitcher value independent of fielders that is as meaningful as ERA for full-time pitchers but directly denominated in net runs.

* * *

“I don’t see the benefit here,” because it ignores IP, but you have to check that there are ‘enough’ IP to trust the results.

Precisely the same thing can be said of any linear weights valuation of a pitcher, or a batter, for that matter.  Pete Palmer has been publishing linear weights runs ‘produced’ by batters and ‘saved’ by pitchers in baseball encyclopedias for over twenty years.  Are all those numbers “toys” as well?

There’s a place for rate stats and a place for net value stats.  If all you know is the rate, you don’t know the impact without knowing the number of opportunities; if all you know is the net impact, you don’t know the rate without knowing the number of opportunities.

* * *

“Now, is SO-2*BB more stable than SO/BB?  Well, that depends on how many innings you have.” I frankly did not understand the logic of what followed. 

* * *

“But, in no way would I use this as anything other than a toy.”

I think many fans will find it useful, again, for the following reasons:

[1] For large samples of innings, it correlates with next-year ERA as well as current-year ERA.  Nobody has ever called ERA a “toy.”

[2] It has approximately the same ‘scale’ as linear weights formulas using HR, bf and/or ip. 

[3] It is the only practical way for a fan watching a game to estimate the net impact, in runs, of the pitcher so far that year, given the information commonly provided at the stadium or on TV.  I’d call that a useful tool, not a toy.

[4] It is a clearly better than ERA for relief pitchers with smaller sample sizes because it focuses only on the items they can control and has the correct implied .500 replacement value threshold.

[5] If the 2009-11 pattern of 2.2 SO per BB persists, you can still do the formula in your head as 2BB, plus one-tenth of 2BB, less SO, divided by three, multiplied by minus one to get runs saved. 

It’s a little analogous to GPA on offense, a terrific stat.  The formula is 1.8 multiplied by OBP plus SLG, divided by 4.  You can easily do (2OBP+SLG)/4 by taking half of SLG, adding OBP, and taking half of that.  If you want to get to the last little bit, subtract another 5% of OBP, which is just one-tenth of half of OBP.

* * *

Kinkaid,

As you can see from the tables above, though “throwing out IP” (or BF) results in some loss of information, as seen when comparing the predictive power of (SO-2BB)/3 with (SO-BB-.085bf)/3, I would not consider the loss “huge.”

Also, the formula should not confuse a fan new to sabermetrics.  You explicitly say to the person that each BB, given bf, is worth about a third of a run, and also that each SO, given bf, is statistically associated with about a third of a run saved, both because of the inherent value of about .27 runs, but also because SO are associated with fewer HR/bf-so-bb and H-HR/bip.  Then all you have to say is that given the current ratio of SO to BB, you have to have more than twice SO than BB to be ahead—given whatever number of batters faced, since both have the same denominator.

Contrast that with FIP, which explicitly assigns “3” runs to BB but only “2” to SO, and for which there is no equally simple arithmetic to explain why that ‘incorrect’ ratio applies.

http://www.insidethebook.com/ee/index.php/site/comments/tangos_lab_deconstructing_fip/

I remember the “deconstructing FIP” thread, which is why I said “no _equally_ simple arithmetic.” Frankly, after a couple of quick reads of “deconstructing FIP,” I still don’t follow the arithmetic, though I’m sure if I sat down with pencil and paper and carefully followed all the steps, I’d get there.

Why the fascination with SO and BB as a sort of natural pair? If I want a quick evaluator of pitcher ability, I’d rather look at SO/HR.

dave,

I think the fascination with SO and BB is that they are the two stats the pitcher has the most control over. 

SO/HR might very well be a great rate stat. 

The point of the above analysis is that we have plenty of good, publicly available rate stats, which appear to have approximately similar power to predict next-year ERA for full-time starters.

Now we have a formula that (1) you can calculate in you head using stats flashed at games, (2) measures pitcher value independent of fielders that is as meaningful as ERA, and (3) is directly denominated in net runs.

Now we have a formula that (1) you can calculate in you head using stats flashed at games, (2) measures pitcher value independent of fielders that is as meaningful as ERA, and (3) is directly denominated in net runs.

And really, what fan hasn’t yearned for such a formula?
:>)

But if someone does, I’d use (K-BB-IP/3)/3.  At least that communicates the right way to think about the relationship: every “net” K is worth 1/3 run, and the average pitcher produces 1 per 3 innings.

Dave, I think the focus on K and BB derives from the idea that both relate to controlling the strikezone.  Also there is a perceived tradeoff between the two, in the sense that hard throwers often have trouble avoiding BBs (similar to tennis players sacrificing velocity/aces on second serve to avoid double faults).

Guy,

Nice finishing touch replacing .085BF with .33IP, which makes (SO-BB-.085BF)/3 formula easy to calculate in your head and gets the same results in terms of the correlations and other stuff already shown.

I now proclaim Guy’s “(SO-BB-IP/3)/3” the greatest formula in the history of sabermetrics for 2001-11 pitchers when all you have is SO, BB, and IP.

It’s two bad Guy that you aren’t Dan Fox, because then I’d call it FoX NPR, for Fox Net Pitcher Runs.

:>)

Going back through history, if you want to flip through an encyclopedia, use Guy’s formula for the following seasons, using the following integer to divide IP (the “Guy Divisor"):

2010-11_______2
1994-09_______3
1984-93_______4

If I have the math right, the Guy Divisor each year is equal to the inverse of (lgSO - lgBB)*(lgBF/lgIP)

Before that, the Guy Divisor bobs up and down between 3 (1963-68) and 6 (late ‘70s and ‘80s), then becomes really unstable for seasons before 1955, because the number of SO and BB were too close together. 

For 1918-1954 I’d use (SO-BB)/2.  As mentioned in Wizardry, SO and BB were both associated with close to +/-.5 runs for years before SO spiked up in the late 1950s.  For 1901-1917 I’d go with (SO-(3/2)*BB)/2, and for 1893-1900 (SO-(2/3)*BB)/2.

---"Dave, I think the focus on K and BB derives from the idea that both relate to controlling the strikezone.”
*******

Sure, I understand that idea. I’ve been around saber for quite a while. But there’s lots more to striking out batters than throwing strikes. And don’t HR relate to “controlling the strike zone”, too? Throw it in the wrong spot and the ball goes over the wall. I simply don’t think that people who see this big natural tie between SO and BB have really thought it through.

44 - But Fox and NPR are two networks on opposite ends!

Brian, that was precisely the joke!

Hey are you still doing work with Gameday batted ball data?

MAH: Colin W. has some nice things to say about your book (but still no review!) here:  http://www.baseballprospectus.com/chat/chat.php?chatId=897.

48 - yes, I use Gameday for play by play, minors and MLB.

Guy, thanks for the link to Colin’s comment.  It’s hard to respond to his claim that his system is better than DRA, because like all BP defensive metrics, it has never been disclosed in a reproducible form.  DRA is also being revised, with the version for 1989-99 and 2003-2011 having been disclosed in the 2012 Hardball Times Annual.  I am currently working on the new 1950-1988 and 2000-02 version which will apply some of Sean Smith’s ideas in a significantly less biased way, as I said I would in Wizardry.  I also don’t believe Colin’s system goes back in time.  I will be very curious to see what he writes in the BP book coming out in a week. 

He’s right about the formatting of the equations.  Oxford literally did nothing and I had to do my best to tell the offshore typesetter exactly what to do, almost character by character.  And Oxford also turned down my request to have some boxed summaries interspersed to help break up the text, which I admit is heavy going in many places.  And the typeface is hard to read.  And I had to do all my own copyediting.  I think you can gather how I felt about the process . . .

Brian, is the Gameday data free or at least open-source?  Where can I get it?

Gameday data is free. This is the top level folder
http://gd2.mlb.com/components/game/

You need some kind of script to download the data. Adler’s “Baseball Hacks” was the first to make one public, written in Perl.

The one that I, and I believe Colin still use is Baseball on a Stick (BBOS) http://sourceforge.net/projects/baseballonastic/ It is a collection of Python scripts that parse Gameday’s xml files on the fly, loading the data in a MySQL database. From those tables I create an optimized data warehouse as the source for analysis.

Mike Fast has Perl code on his blog. I know Mike’s downloads all the original files, then starts parsing. Max Marchi recently created R scripts which I believe do a parse & save.