THE BOOK--Playing The Percentages In Baseball

Hey Tom,

So the break-even SB% is ~ 71%?  Would I be ok using that as a quick+dirty method for players and teams?

That includes all SB (3B, home).

The breakeven for 1999-2002 of 2B is .687, as detailed in The Book.  That’s for a 5.0 runs per game environment.

For 4.7, it’s probably something like .675.

Best thing is not to drive yourself crazy, and presume .6667.

Tom

The empirical run-values I have for the SB and CS from 1999-2002 are:

SB: 0.168
CS: -.450

There are a few reasons why my values differ from Tango’s:

1. Tango treats double steals as one stolen-base. Retrosheet codes double-steals and triple-steals as one event. I treat double-steals as 2 seperate events. Therefore, my run-values for the SB will be slightly lower than his.

2. Tango’s run-values don’t include SB and CS that occur on strikeouts or walks. Retrosheet will code a K+CS, as a strikeout. In my linear weights, I split the strikeout/walk portion of the event from the basestealing portion of the event. I always treat the strikeout or walk as the 1st event, and the basestealing event as the 2nd event. The run-values of these stolen bases and caught stealings where the plate-appearance is terminated are significantly lower than stolen bases and caught stealings that occur DURING a plate appearance. Therefore, my run-values for stolen bases and caught stealings will be slightly lower than Tango’s.

3. Tango’s run-value for caught stealing doesn’t include pickoff caught-stealings, since Retrosheet codes them as pickoffs.

I may be wrong, but it seems to me that if one had to make a simplistic choice, that counting all pickoffs as CS, would lead to greater accuracy than counting none of them.

terps/3 - I recently ran queries for po & pcs when I was looking at steals. I’ve published the results for catchers at FanGraphs, pitcher’s article is coming, which will focus on pickoffs. You are right that Retro incorrectly codes pcs before 1999, and I pointed this out on their mailign list. The play description says either pick off or caught stealing, but not both - but by querying the putouts and assists when either is true I was able to distinguish cs (by catcher) from po at 1b and pcs at 2b. Steve Carlton led with I believe 319 total pickoffs.

Question for terps and Tango - when calculating the run value of the caught stealing, which base/out states do you compare? Consider this, and tell me if you are doing it this way or not, and if I make sense or not - I was trying to conceptualize the difference between a caught stealing and a pickoff. If the defense fails in a pick off attempt, there’s a runner at 1b. However, if the defense fails in a caught stealing attempt, there’s a runner at 2b. Therefor, the penalty for a cs should be a runner removed from 2b, which is higher than the penalty for being picked off, which is a runner removed from 1b.

Brian,

Using your example, from the perspective of the offense, I would assign the same run-value for the pickoff as I would the caught-stealing. The penalty for getting caught-stealing second-base is the removal of the runner from 1st base. Because the average pickoff (not counting pickoff errors where neither the catcher or pitcher are credited with an assist) occurs in higher leverage situations than the average caught-stealing, the run-value for a pickoff is slightly higher than a caught-stealing.

I think what you are after Brian, is the run-value of the pickoff and caught-stealing from the defense’s perspective. What do you mean by failed pickoff-attempt. You said: “If the defense fails in a pick off attempt, there’s a runner at first base.” What about pickoff-errors, where the runner advances one or more bases?

Brian,

Your pickoff numbers for Carlton look nothing like mine. Here are the base-stealing stats I have for Carlton:

92 Pickoff CS
53 Pickoffs
15 Pickoff Errors
314 Stolen Bases Allowed
133 CS (not including Pickoff CS)

How are you getting 319 pickoffs for Carlton?

"that counting all pickoffs as CS, would lead to greater accuracy than counting none of them. “

Who said I count none of them?  I count every event, but I simply keep it separate.

PO% Opp SB CS PC PK PE
Carlton, Steve 2b .059 5146 240 91 170 136 13
Carlton, Steve 3b .004 3352 70 14 9 4 2

Pickoff error is a seperate category, and gets it’s own run value - how many bases did the target runner advance, how many bases did any other runners advance, on average

You’ve got 314-133 on throws from catcher, I’ve got 310-105. I did not count any attempts from 3rd to HP. You’ve got 92 PCS, I have 179, you have 53 PO, I have 140. We agree on PE.

On pct of runners in stealing situation (runner on 1st, none on 2nd), Carlton by my numbers is below several others. 500 or more opps, Beimel has .113 (58 of 515). The league wide numbers for olders seasons are consistent with 1999-2008.

Here’s my code

Sum(IIf([Events.START_BASES_CD]=1
Or [Events.START_BASES_CD]=5,
1,0)) AS Opp,

Sum(IIf([Events.RUN1_SB_FL]=’T’,
1,0)) AS SB,

Sum(IIf([Events.RUN1_CS_FL]=’T’
And [Events.ASS1_FLD_CD]=2,
1,0)) AS CS,

Sum(IIf([Events.RUN1_CS_FL]=’T’
And [Events.ASS1_FLD_CD]<=1,1,
IIf([Events.RUN1_PK_FL]=’T’
And [Events.ASS1_FLD_CD]<=1,
1,0))) AS PC1,

Sum(IIf([Events.RUN1_PK_FL]=’T’
And [Events.ASS1_FLD_CD]=1,
1,0)) AS PK1,

Sum(IIf([Events.RUN1_PK_FL]=’T’
And [Events.ERR1_FLD_CD]=1,
1,0)) AS PE1,

Sum(IIf([Events.RUN1_SB_FL]=’T’
And [Events.ERR1_FLD_CD]=2,
1,0)) AS TE

Brian - If I am reading your query for PC1 correctly you are double counting the PK1s by also including them in the PC1 totals. With your nested IIf statements PC1 will record as 1 if either if statement is true.

I caught that myself when rereading the post. I checked my totals and everyone had a total PC >+ PK. I had written a reply but it vanished into etherspace. I changed the PC test to if PK=T and ASS2=3, and am rerunning now.

Rebuilt the data warehouse tables, now have Carlton with 109 PC, 160 PK, still career leaders in each.

By failed pickoff attempt, a throw to first where the runner is safe. The defense made an attempt to put out the runner, which failed, the runner is at first.

When the runner takes off for second, the throw arrives at 2b, that’s where the tag is made. If the call is safe (a failed attempt by the defense) then there is a runner at 2b, therefor if the runner is called out, I believe it should be a runner taken from 2b, and that’s what the out value should be based on. If it’s true for defense, it’s the same for offense, which is the flip side of the same coin.

If you don’t value it this way defensively, the pickoff at 1b has the same run value as cs at 2b, and I don’t think that’s right.

My guess would be that the % of successful steals when the pitcher goes to 1st is much lower than when he goes home.  Obviously this can be checked.  If this is true, I think the fundamentally different plays are those that the pitcher throws to first versus home, not depending on which way the runner then turns.  So, if throw to 1st, safe at 2nd is as rare as I think, I would rate the run value of the out the same at either base after the pick-off.

If you don’t value it this way defensively, the pickoff at 1b has the same run value as cs at 2b, and I don’t think that’s right.

I don’t know why a pickoff at 1B shouldn’t have the same value as a caught stealing at 2nd.  Both plays start with a man on first and end with no one on.  Why shouldn’t they be worth the same?  Just because one event was initiated by the runner and the other initiated by the pitcher shouldn’t make a difference.

Because a failed PO attempt by the defense results in a runner at 1b, while a failed CS attempt by the defense results in a runner at 2b...the tags are being made at different bases

The value of a pickoff or CS shouldn’t depend on the value of a “failed” PO or CS. They are different events. I agree with Peter J.

I agree with the majority. 

You have to think in terms of what the run expectancy was before and after the play.  The difference is the value of the event.  Whether the catcher throws the runner, whose current earned state is at 1B, out at 1B or 2B is irrelevant.

You have to think in terms of what the run expectancy was before and after the play.

One exception to this might be when there is a compound event, such a successful steal of second with an overthrow and the runner tries to go to third and is thrown out.  This is recorded as an SB by retrosheet, and the run value is usually calculated as a negative value equal to the removal of a runner from 1st.  This is mainly for expediancy.  You could make the argument that it should instead be split into 2 events. The first a normal steal of 2nd with a run value equal to the change of moving a runner from 1st to 2nd.  And a second event which was a base running error by the runner, with a run value equal to the removal of a runner from second plus and out.

Peter - yes, I would split that into two events, as I believe scoring rules would dictate.

Well, that idea’s been shot down, I just wanted to make sure everyone knew exactly what I was arguing.

Next question...lhp pick a runner off 1b at a rate of 2.4%, while for rhp it’s 0.9%, just an inherent difference in being right or left handed on the mound...so, when I am normalizing each individual pitcher’s po% to league average, should I compare them to the overall mean of 1.3%, or compare them to what the pitchers of like hand do? Dwight Gooden and Ron Guidry both have a career po% of 2.4%, which for Gooden is 2.7x avg and 14th highest (1000 opps) for a rhp, but for Guidry is exactly what’s expected of a lhp. I lean towards norming them against all pitchers, regardless of hand, but I wanted to get the group’s opinion.

Certainly not to the league average as a whole.  By handedness is the way to go and I would use Z score instead of % of average.

"One exception to this might be when there is a compound event, such a successful steal of second with an overthrow and the runner tries to go to third and is thrown out. “

It’s irrelevant to Brian’s argument, as he originally stated.  If you want to give +.18 runs for the SB and then -.65 for the CS at 3B, or whether you give him -.47 for the CS and zero for the SB, it’s the same thing. 

Now, you would prefer to give it as two things, since two things happened.  And, so, in an accounting sense, you have .18 in the positive ledger and .65 in the negative.  But, if someone wanted to call it 0 in the positive and .47 in the negative, I wouldn’t have much objection to it.

What Brian was originally arguing though was wrong, because the accounting wouldn’t have added up.  If the runner starts at 1B, then it is irrelevant if the catcher gets him out at 1B or 2B, since the runner never got a plus in the positive ledger to begin with.  Without a plus in the positive ledger, the accounting is all on the negative side.  And therefore, it doesn’t matter where the runner was putout.

It all comes back to the run expectancy matrix.  As long as it adds up, then we can each present reasonable arguments.

It’s irrelevant to Brian’s argument, as he originally stated.  If you want to give +.18 runs for the SB and then -.65 for the CS at 3B, or whether you give him -.47 for the CS and zero for the SB, it’s the same thing.

That’s not what I said, not what I would do, and not what I would recommend doing.  Nor do I say I would prefer to see it as two events.  I said that one “could make the argument” to do that.  But not as a SB and a separate CS. Never.  The attempt to go from second to third should be lumped into a separate “other baserunning errors” category along with being out trying to take an extra base, out trying to advance on a wild pitch or passed ball, etc.

The times league avg was just for illustration, z-scores are fine, plus a normalized percentage.

"The attempt to go from second to third should be lumped into a separate “other baserunning errors” category”

Yes, agreed.  I spoke too fast in saying “CS”.  Otherwise, what I said stands.

So, giving +.18 for the SB and -.65 for the “Out advancing” or -.47 for “baserunner movement”, it’s the same thing.  As I said, I’d keep it as 2 events, not one.

The larger point is to prove why what Brian was saying was incorrect, about having the player be thrown out going back to 1B as opposed to moving forward to 2B.

Brian,

I have not followed this thread, but you always want to compare a player to all players (RH and LH) unless you already have a LH/RH adjustment for handedness.  Where the positive or negative value comes from (their handedness or their actual “skill") is irrelevant.  Now, it is also nice to let readers know about the inherent difference due to handedness, but is it not necessary.  Do we compare all tall players to other tall players in offensive metrics?  Etc.

An example of where you HAVE TO compare to similar-handed players is when I do my offensive lwts.  First I use different lwts values for RH and LH ground balls, and these values include the GDP.  So when I do the GDP addition, I have to compare a RH batter’s DP rate to other RH batters, since I already account for the differences in DP rate in the ground ball lwts values.

I don’t have to do it that way.  I can account only for a runner advance from second base on an out, and first to third on a hit (lefties have an advantage in both), but NOT the GDP.  If I do it that way, then when I add in the GDP value, I would compare everyone to all other batters, regardless of handedness.

I hope that makes sense.

MGL, that does make sense, and I said I was leaning towards that position, but I just wasn’t sure and wanted some input.

If a hitters bats left, he will on average be quicker down the line than someone who bats right. This effects infield hits, roe, DPs, etc. I recently started writing queries for gb DPs, and I held for type of turf and batter handedness, doing four different calcs and then combining.

In this case, being a lhp immediately gives the pitcher a advantage in pickoffs. How much of the result is the skill of the pitcher, and how much because of the hand he throws with? Or should they just be combined?

Tangentially related:  According to the B-Ref league splits, there were only 66 SBAs on the first pitch last year, while in prior seasons there were typically more than 900 first pitch SBAs.  Seems highly unlikely teams would make such a radical change in one year.  Anyone know if this reflects an error in Retrosheet data (or at B-Ref), or maybe some change in methodology (like SBAs are now linked to the count following the SBA, rather than prior)?

Guy, great point, and I’ve sent your post to Sean.

It was a bug in in-season update program.  I build my event files in retrosheet form by hand from the pbp format sent to me by the provider.  For baserunning plays, I was mistakenly including the count after the play rather than the count before the play.  I actually just noticed it myself this week as I’ve been incorporating the new retrosheet data into my db.