THE BOOK--Playing The Percentages In Baseball

Yeah, I can see how spreadsheets would negatively impact office productivity.

(My gut feeling is that it’s actually a concern someone would put proprietary company stuff on a public server. I just liked the wisecrack.)

As someone who uses offensive PADJ (while acknowledging the weaknesses of doing it that way), I will not criticize Chris for that.  But I believe he uses one-year (maybe even one-league) adjustments, and that is just not

I think that approach has utility for comparing teams (i.e. the AL shortstops couldn’t hit the side of a barn this year, so Jhonny Peralta was a big relative plus for the Indians), but for valuing the players, it’s a disaster (Peralta is not a star simply because his 13 peers were hapless).

I think that at the very least you need to look at all players in that position in both leagues, and adjust for league differences seperately. If all the good shortstops are in the NL, it doesn’t make a mediocre shortstop in the AL more valuable than a mediocre shortstop in the NL (once you’ve controlled for the league quality difference, of course.) I think BP does it the wrong way for VORP.

Patriot, is that correct?  Yowza, I feel bad for the NL LF of 2001-2004, who must have been an extra minus 7 or 8 runs each, just to compete against Barry Bonds:
http://www.fangraphs.com/statss.aspx?playerid=1109&position=OF

I see no reason to separate NL from AL, and even less of a reason to separate LF from RF.

Can’t Chris, and I, and MGL, and Patriot, and the BP guys, and everyone else here have it out, once and for all, as to the idea that the NL LF must not necessarily be equal to the AL RF, each and every single year?

You’d have to ask Chris, of course, but that is my recollection.  Here is a comment (#27, by Chris) from a BTF thread his 2007 figures:

I went over these again, and if there is a calc error (and there still could be), I cannot find it. I also put another run estimator - Tech RC in the spreadsheet and calculated the differences. THere were some players off by as much as 15 runs, but it was speedsters (Juan Pierre, Rickie Weeks), and usually higher in my work. Carlos Lee has 110 XR, in 477 Outs. All NL LFs generated 0.2218 runs per out. They averaged 108 RC over a season of outs (487 per team, sum of all LF outs divided by 16).

The HOU PF is 100. So Lee generates 3.8 XR+ accounting for the slight difference in outs. The entire NL generates 0.19394 x-runs per out. In Carlos’ outs (477), the average NL hitter would generate 92.5 x-runs. So, compared to NL average, Carlos is +17.1.

Carlos hit into 27 GiDPs, so anything not accounting for that would increase this.

#35:

Dial quoting previous post: “Comparing players only to their own league might explain a lot of the apparent strangeness, though I’m not sure that’s the best method given interleague play.”

Dial: My personal philosophy in this is that when it comes to “competing”, the NL teams only compete with the other NL teams wrt making the playoffs.

Perhaps my impression is off, but it sure seems to me as if he is just using one league.

Here’s the only thing I could find addressing the issue for VORP:

http://groups.google.com/group/rec.sport.baseball.analysis/browse_thread/thread/82b52bab71bb0039/bf9b414686dd2890?lnk=st&q=vorp+league#bf9b414686dd2890

So at one point Woolner looked at league averages per position, and reconsidered the position. I obviously can’t be certain about how they’re handling the issue currently. I also don’t know how Davenport handles the issue for WARP.

Tango,

If you log onto a proxy server, you could probably bypass your internet censor.

If Chris wants to hold with the view that “my 1B is better than the 1B in the rest of the conference”, that’s fine.  What he cannot do is then take Rollins relative to the other NL SS and Howard relative to the other NL 1B and presume you can now directly compare the two.  The average NL SS is not necessarily paid or is worth exactly the same as the NL 1B, and when you merge these lists, that’s exactly the implication being made.

Generally, I do not enjoy criticizing other people’s work. But you guys have already pointed out 2 big flaws in Chris’s methods.

The first problem is that he uses XR. The full version of XR overvalues singles, undervalues doubles, overvalues HR, overvalues IBB, overvalues GIDP, undervalues SH, and overvalues SF. It is rediculous to weight a SF as being 6 times more valuable than a regular batting out. In XR, the SF has a run value of .37, the out has a value of -.09. Patriot has a full version of ERP on his website. Chris would of been better off using that. 

The second problem is that he uses 1-year positional adjustments. Actually, his positional adjustment is implied because he compares a players RAA to the league average player at a specific position. Rather than doing this, he should of calculated the RAA first, then applied Patriot’s positional adjustments that he has made available on his website. Patriot’s positional adjustments are based on multiple seasons worth of data.

I think the criticism is fair, if the author doesn’t note all the implications of the decisions being made.  For example, if Keith Woolner would have made clear that the walk is severely undervalued, then he wouldn’t need me to point that out.  That all of BPro didn’t know this makes my posts on the topic more than legitimate.

I didn’t know that Chris was using positional adjustment based on the average for the year/league in question.  Patriot points it out, so it is more than a fair thing to criticize.

Have you ever used XR Tango? No. I never recall a single instance where you have ever used it. Absolute LW, TT Baseruns, ERP are all better run estimators. There are a lot of smart people who still continue to use bad run estimators.

I stated earlier that XR over-values the GIDP. This was wrong. I forgot to add the batting out to the value of the GIDP.

You probably shouldn’t try using a proxy server Tango. I don’t want to get you fired. It was just a trick that I used to bypass the internet censor when I was in high school.

Here are the totals:
Pos NL AL Tot
1B -2 -11 -14
2B 13 11 23
3B -3 1 -2
C 1 5 6
CF 6 7 13
LF -21 2 -19
RF 11 14 25
SS -4 13 9
DH 0 -11 -11
OPD 0 30 30

Obviously, he is not zeroing out.  He might be putting players at their primary position.

Are they supposed to zero-out by league, or zero-out combining both leagues? If his data was grouped by league, then he just made a mistake somewhere in his AL calculations.

OK, he did group his numbers by League. The DH field for the Angels is a null value on the AL Team OPD sheet. This could be the culprit.

If we are giving players credit for a SF, then we might as well just use runs and RBI.  Or why not a SGB (a ground ball that scores a run).  To have SF as a term with a positive coefficient in a “run estimator” that is otherwise context-neutral just boggles my mind.  A SF is an out in which a run happens to score.  Why is that (as opposed to just another fly out) in a linear weights formula, like XR?  If your answer is because it is a run estimator formula and it makes the formula correlate better with actual runs scored, then I ask again, why is a fly ball that scores a run singled out, as opposed to a ground ball that scores a run or a single that scores a run (or two) or a HR that scored 3 runs, etc.  If your answer is that, “It is the only thing in a traditional stat line that tells us when a run is actually scored,” then my next comment is, “Is your formula trying to capture a skill in the player that might have some predictive value if that player were to move to another team?” If yes, then SF has no business being in that formula.  If you are just trying to capture “run making performance” with no regard to whether that player had anything to do with some of that run making, then, as I said, to use a SF in the middle of bunch of other completely context-neutral terms is so ludicrous, it makes me laugh.  And, as many people have already said, why is ANYONE using anything but linear weights?

I imagine that Chris uses XR because Baseball Think Factory is Jim Furtado’s site. Knowing what we now about Linear Weights and Run Estimators, there is no excuse to use a regression-based formula like XR. A lot of sabermatricians seem to be stuck in 1999, when Tangotiger, MGL, David Smyth, and Patriot weren’t around to educate sabermatricians on Linear Weights and Run Estimators. You guys have been championing Linear Weights and Baseruns for almost a decade. Then, why do Bill James and Sean Foreman continue to use Runs Created? Why do Furtado and Dial continue to use XR? Bill James’ website even has Run Expectancy tables and yet we don’t see any Linear Weights. We get the same old Runs Created, even though someone with half-a-brain could generate absolute Linear Weights from the RE tables on the website in a matter of a couple of hours. It is frustrating to see very intelligent sabermatricians not doing their homework on run-estimators.

As far as the SF goes, I couldn’t agree with you more MGL. I do keep track of the SF when I calculate Linear Weights. However, this is only so I can seperate the ROE SF’s from SF Outs. I then count the ROE SF’s as plain-old ROE’s, and count the SF’s as plain-old Batting Outs. I do include normalized GIDP’s in my Linear Weights. In fact, if it wasn’t for one of your old Fanhome posts, I wouldn’t of known how to normalize GIDP’s. Brian Cartwright helped me with this as well.

We can probably blame Bill James for popularizing the Sac Fly and GIDP since he decided to include these categories in his technical Runs Created formula 20+ years ago. Pete Palmer, recognized the situational nature of these categories 30 years ago, and made the correct decision by not including them in his Batting Runs formula. The more I study sabermetrics, the more I begin to appreciate Pete Palmer. The guy is a genius.

Noticed a bunch of typos in my last post:

2nd sentence, 1st paragraph should read “Knowing what we now know ...”

3rd sentence, 2nd paragraph should read “I then count the ROE SF’s as plain-old ROE’s, and count the SF Outs as plain-old Batting Outs.”

2nd to last sentence, 3rd paragraph should read “The more I study sabermetrics, the more I appreciate Pete Palmer.”

Sorry for the typos.

"Then, why do Bill James and Sean Foreman continue to use Runs Created?”

Foreman has RC on his player pages, but also has Pete Palmer’s batting runs.

Right.  I don’t mind that RC is shown, as long as something better is also shown.

I object to OPS+, because something better could be shown, but isn’t.

Referring to post Tango’s post #8, what is the correct way to compare players from different positions?

If you are just measuring offensive value, can you use RC/RAA to compare players from different positions after making the positional adjustment. Would it be fair to compare Mauer’s position-adjusted RAA to Youkilis’ position-adjusted RAA?

For defense, I guess that you can only compare a player’s defensive RAA to other player’s who play that position.

You compare hitting to ALL hitters, not just those at his “position”, as if “position” is something fixed. 

1. Compare offense to all offense
2. Compare fielders to fielders at their position, and then add a positional adjustment so that cross-position comparisons are possible
3. Adjust for league
4. Include a playing time adjustment

That’s it.  It’s that easy.

"Compare fielders to fielders at their position, and then add a positonal adjustment so that cross-position comparisons are possible.”

What kind of positional adjustment?

Offensive Positional Adjustment?
Defensive Positional Adjustment?
Offensive + Defensive Positional Adjustment?

To reiterate what Tango is saying and perhaps make it even clearer, if you want to compare 2 (or more) players, simply compare their offense (and baserunning) to any baseline you want and then imagine they play the same fielding position.  In order to the latter, you simply do a positional adjustment to one or the other (or both - it doesn’t matter).

That is the same thing as #22 above.  It is real simple, as Tango says.

If player A is 10 runs above whatever baseline you want for hitting, and so is player B, and player A is -6 at SS and player B is +3 in LF, you simply ask, “What would player A do if he played LF, given that he was -6 at SS?”

Or, “What would player B do if he played SS, given that he was +3 in LF?”

The answer to each question is the same of course.

Now, the way we answer those questions (with positional adjustments) is arguably not perfect, but we do it the best way we can.

We can certainly say the same thing about any other adjustment we do to compare players - park adjustments, strength of opponents adjustments, etc.  None of them is perfect, but we do them anyway to the best of our abilities.

To be less specific about what MGL said, I would ask: “How would Willie Bloomquist or Melvin Mora do if he played SS and LF?”

THAT is the common baseline.  Someone will always say that Jeter may be a worse SS than, say, I dunno, David Eckstein, but that Jeter would be a better LF.  So, instead of trying to deal with specific players playing out of position, and all the implications of that, I sidestep the discussion by talking about versatile players like WFB and Mora. 

And so, if Jeter is -10 at SS and WFB is -7.5, and if Crawford is +9 in LF and WFB is +7.5, then the fielding+position baseline (aka, Willie Bloomquist line) makes Jeter as -2.5, and Crawford as +1.5.  So, the gap in their fielding contributions to their team is effectively 4 runs.

"The answer to each question is the same of course.”

It does get complicated though in that different positions field different numbers of chances.

For my outfield projections, I have CF as 10 runs better than left per 150 games, in the same # of chances (can’t remember if I use the CF average or the corner).  Figure that, and then figure the player projection at that position, with a CF getting 30% or so more chances.

The result is that a guy like Endy Chavez is +19 in center, +22 a corner (or similar) but a bad fielder is -15 in a corner, -30 in center.

So is Endy 49 runs better than Adam “ManRam” Burrell? or just 37 runs better?

Rally,

How are you getting 49 runs for Endy compared to Burrell/Ramirez/Dunn? Can you explain your calculation.

If Endy was +19 as a centerfielder, and Dunn/Burrell was -30 as a centerfielder (based on the position difference and having more chances to field).

As left fielders they are +22 and -15 - fielding skill remains the same but there are fewer chances for a left fielder to make/miss plays.

Rally is right.

Suppose that Willie Bloomquist makes 0.80 outs per play in CF and 0.75 outs per play in LF (in this illustration, it’s harder to make a play in LF).

Let’s say that the average CF has 4 plays per game, and the average LF has 3.  Let’s also presume that the average CF makes .805 outs per play, and the average LF makes .735 outs per play.

So, the average CF is +.005 above Willie, per play, and the average LF is -.015 compared to Willie.

With 4*162 plays in CF, the average CF is +3.2 outs.  With 3*162 plays in CF, the average LF is -7.3 outs.  All compared to Willie.

Now, say you have Endy, who makes +.05 more plays than Willie. 
- In LF, he’d be .05*3*162= +24 compared to Willie and +31.3 compared to the average LF
- In CF, he’d be .05*4*162= +32 compared to Willie and +29 compared to the average CF

On the other hand, if you had Dunn who makes -.05 plays compared to Willie:
- In LF, he’s -24 compared to Willie, or -16.7 compared to the average LF
- In CF, he’s -32 compared to Willie, or -35.2 compared to the average CF

So, Endy “looks” like he is +30 compared to the average at CF or LF, but that’s a trick based on number of opps.

Dunn “looks” like he gets a 19 play adjustment between LF and CF, but again, that’s based on leveraging opps.

I guess you either take the average of the 37 and 49 or you give each player the benefit of the doubt and use the position that best leverages their skills or lack thereof (and use the 37?).  I don’t know.  I guess it is not as easy as I thought.  I have to think about it.

Yes, the one that best leverages them.  So, Endy is +32 compared to Willie in CF, and Dunn is -24 compared to Willie in LF.

This is the same like Mariano Rivera: because of his talent, he gets to have more impact in close games, so he gets the closer bonus.

Endy gets the CF bonus.

This also applies to hitters.  A good hitter will have more than 4.3 PA per game, and a bad hitter will have less.

Well, sorry I missed this.
1) I believe the leagues wrt value should be seperated. And I do so.  I haven’t seen a reason not to for the purposes I do the analysis.
2) I use XR because I started using it 10 years ago (about) and that’s how my spreadsheets are already set up.  I worked to convert to BseRuns, but it doesn’t work.  So provide me a differnt run formula, and I don’t mind using it.  For all the “XR sucks” complaints, you are quibbling over a run or less.
3) I absolutely zero out the leagues.  It’s the “primary position” that makes it appear not so.  The math is correct.
4) Positional adjustments.  I disagree with that philosophy.  People are at the plate at a position.  There isn’t a position of “hitter”.  You hit as a position player (a 1B, a 2B, whathave you).  VORP does this with better precision than I do, but a Q&D analysis indicates this difference across a league is next to nil, and few players would be impacted at all.
What he cannot do is then take Rollins relative to the other NL SS and Howard relative to the other NL 1B and presume you can now directly compare the two.  The average NL SS is not necessarily paid or is worth exactly the same as the NL 1B, and when you merge these lists, that’s exactly the implication being made.” No such implication is made anywhere in anything I have written.  I rarely (never) speak of salaries.  However, the comparison is *still* completely legitimate, and probably moreso if you subsequently made your positional adjustments on salary, because that isn’t affected by “the human factor”.  Jeter may or may not be suited for CF.  ARod may or may not move from a +9 SS to become a +15 3B, or he can move to 3B and become a =5 3B.  His performance is what it is - his salary depends on what someone will pay him, so I believe there’s nothing wrong with my cross-positional comparisons.

What haven’t I addressed? 

Oh, all these are fair criticisms.  I appreciate the feedback to try to improve any system I produce.

"So provide me a differnt run formula, and I don’t mind using it.”

Sure.

A Linear Weights formula would be perfect for the offensive component of your total player ratings. It would save you the hassle of having to convert from absolute runs to runs above average.

Now, I don’t know how many seasons you would want the Linear Weights formula to encompass. You tell me. 1993-2007? 1986-2007? 2001-2007?

I can guess that what you can provide Chris the easiest is the XR “form”, but with the better weights.  So, keep the same categories, but just give him the correct weights.  That’ll be the most painless solution. 

XR exists after ERP and LWTS came on the scene.  If the difference was only one run, then why did XR come to exist?  I would think the difference would be more, especially depending on how the SF is treated.

Even so, we come from the school of 18 different tech versions of RC each barely better than the other.  We’re looking for accuracy!

***

“There isn’t a position of “hitter”.  You hit as a position player (a 1B, a 2B, whathave you).”

Chris and I def have a philosophical difference here.

***

“No such implication is made anywhere in anything I have written.”

Is the average player at each position in the league exactly equal to zero?  If so, then this is what I’m talking about.

Chris,

Here are two formulas that you can use in place of XR. The first is a Runs Created formula and the second is a Linear Weights formula. They have the exact same structure as the original XR formula. I even rounded the coefficients to the same number of decimal places as the original formula. The coefficients were derived from empirical Linear Weights from 1986-2007, a run-context of 4.67 runs-per-game. Let me know if you would like me to use a different set of seasons. A more recent set of seasons would definitely work better for your OPD metric. As usual, the RC version works better than the LW version. This is due to the fact that Linear Weights has a wider range of values for the out categories than Runs Created does.

XR RC = (.47 x 1B) + (.78 x 2B) + (1.06 x 3B) + (1.40 x HR) + (.33 x (HP+TBB-IBB)) +(.17 x IBB)+ (.18 x SB) + (-.28 x CS) + (-.074 x (AB - H - K)) + (-.112 x K)+ (-.43 x GIDP) + (.14 x SF) + (.10 x SH)

XR LW = (.47 x 1B) + (.78 x 2B) + (1.06 x 3B) + (1.40 x HR) + (.33 x (HP+TBB-IBB)) +(.17 x IBB)+ (.18 x SB) + (-.44 x CS) + (-.247 x (AB - H - K)) + (-.285 x K)+ (-.59 x GIDP) + (-.05 x SF) + (-.12 x SH)

A 0.038 run difference between a batting out and a K?  Does the batting out include or exclude the GIDP?

Tango,
we do have two philosophical differences around that.  and they both involve hitters and then the defensive positional adjustments.  And that’s fine.  I’d just prefer you not say mine is wrong, but rather you don’t like to do it that way.

Each position is zero for average.  I don’t know why that means what you say it means.  It is how many runs that a player is outproducing his peers.  The Phillies need to outpeform the other NL teams to win games, and they have to outperform them with fielders hitting.  I think it works.

I’ll have to check, but I think the run values in #37 are the same as mine.

I don’t have a difference of 0.22 - I am pretty sure it is 0.31.

Also, when people convert these plays to runs (like your “runs at position”, Tango), what are you using to determine runs per play?

Chris, I thought I did characterize it as a somewhat philosophical issue. You are true to your philosophy, so you are not “wrong” on that end, unlike say the math errors that other saberists do.

However, I do think it is wrong to include Barry Bonds in the 2001-2004 NL LF pool, and is completely ignored in the NL RF pool, and the AL corner OF pool, when it comes time for comparison purposes.

Brian Giles was a LF in 2003 and a RF in 2004.  I don’t see why he gets to be compared to Bonds one year and not the other. 

So, it is not totally a philosophical issue, just somewhat.

***

As for the .22/.31 difference, maybe it would be more instructive to print out the actual equation you are using if it’s not the last published version of XR, which we’ve got here:
http://tangotiger.net/wiki/index.php?title=Extrapolated_Runs

"Also, when people convert these plays to runs (like your “runs at position”, Tango), what are you using to determine runs per play? “

Actually, I am using UZR runs. 

***

However, in terms of converting plays to runs in a general sense, I would think .75 for middle infielders, .80 for corner infielders and .85 for outfielders would be a close approximation to more rigid numbers, which I know you’ve published in the past.

That said, if someone is +20 plays, which is a fairly extreme number, the above numbers would convert to +16 runs +/- 1 run.  So, in terms of making a quick explanation, I just use the .80.  But, if we want to do a better job, then a realization that extra base hits have much more of an impact in the OF than the middle infield is in order.

Here’s the XR Reduced formula, from the wiki:

XR Reduced = .50*S + .72*D + 1.04*T + 1.44*HR + .33*(W + HB) + .18*SB - .32*CS - .098*(AB - H)

Compare to:

XR Reduced = .46*S + .76*D + 1.03*T + 1.41*HR + .29*(W + HB) + .20*SB - .26*CS - .08*(AB - H)

Which are the weights I come up with using an empiric apporach (including some adjustments to make everything add up properly due to missing events, and reconciling per out).

I used data from 1955 through 1997 (the exact same data set as Furtado used in XR) so this is an apples-to-apples comparison. I’m just trying to show the differences in approach here.

Now let’s reconcile by PA instead:

XR Reduced = .61*S + .91*D + 1.19*T + 1.56*HR + .45*(W + HB) + .20*SB - .40*CS - .14*(AB - H)

For a lot of reasons, I would prefer the third formula, especially if you’re measuring playing time by plate appearances (and let’s face it, when you’re presenting offensive stats to a broad audience, you’re going to have a hard time getting people to understand outs as a denominator.)

I can break out the different categories of walks, I can break out strikeouts from outs in play - I dislike breaking out SF/GIDP because then you’re mixing context-neutral and context-dependent events and it gets a bit hairy. I can also look at a more appropriate runs context; for modern offensive players I find 1994-2007 works well.

Good job Colin, in first presenting it as an apples-to-apples comparison.

I will get this and publish it in that and future articles.

"A 0.038 run difference between a batting out and a K?  Does the batting out include or exclude the GIDP?”

AB-H-K includes the run value of ROE’s.

GIDP is -.074 + - .43 = -.50. So, the batting out is included for the GIDP under AB-H-K, but it is weighted at the wrong value. In reality, the batting out for the GIDP is usually around -.24 runs.

Why make an apples-to-apples comparison to XR? What’s the point? XR uses too wide a data range (1955 to 1997). The AL scored 5.4 RPG in 1996 and 3.4 RPG in 1968. With that in mind, I figured 1986-2007 was too big of a sample. So I used 2001-2007 instead (4.74 RPG). The values don’t change that much. Chris only uses OPD for recent seasons, so I thought these formulas would be more useful. RC1 is reconciled using the PA approach that we have been discussing, but instead of using PA I used PA+SB+CS. Instead of posting 3 seperate equations, I will just post a table with the 3 sets of values. These values would get plugged into the XR formula.

RC, RC1, LW, EVENT_NAME
0.47 0.59 0.47 1B
0.78 0.90 0.78 2B
1.06 1.18 1.06 3B
1.40 1.51 1.40 HR
0.33 0.45 0.33 HP+TBB-IBB
0.17 0.29 0.17 IBB
0.18 0.30 0.18 SB
-0.27 -0.32 -0.44 CS
-0.075 -0.134 -0.251 AB-H-K
-0.113 -0.171 -0.289 K
-0.43 -0.48 -0.60 GIDP
0.09 0.00 -0.12 SH
0.13 0.05 -0.07 SF

You’ll probably be happy to here that I’m taking a break from all this run-estimation stuff. I’ve been examining it in some form every day over the past 4 months.

The sole purpose of the apples-to-apples is to highlight the differences in methodology (team-level versus game-level) and not the dataset.  Once one is satisfied that the game-level analysis is better than the team/seasonal-level analysis, THEN we can move to the next step and request the game-level LWTS or RC for a particular run environment.  I like to use 1993- or 1994-present, since there’s such a clear change in runs right around there, and it keeps me from trying to figure if I want to start in 1997 or 2001 or 2004 or whatnot.  1993-present is stable, and I get a huge sample size, so that’s my general recommendation.

I’ve tested run formulas from 1993-2007. They over-estimate runs scored for the most recent seasons.

RPG 1993-2000: 4.90
RPG 2001-2007: 4.74

Although, I won’t argue with using 1993-2007.

Here is how I grouped LW to set up seperate Baseruns equations for each league to generate single season LW for each league:

1954-1962
1963-1972
1973-1976
1977-1985
1986-1992
1993-2000
2001-2007

That’s 14 seperate Baseruns equations, one for each league. I then used these equations to generate single season LW and RC. Full Baseruns equations were used, although I didn’t end up using all of the categories in my final LW and RC formulas.

The RPG from 1986-1992 was exactly the same as 1977-1985, but HR increased 28% and SO increased 10% compared compared to 1977-1985.

I promised that I was going to take a break from this run-estimator stuff. Help! I am obsessed with run-estimators.

Of course, I used RC Linear Weights for the Baseruns equations. I converted them to Linear Weights by using an “R/O rate” for each event that was derived from each dataset (Event R/O / Total R/O).

I was happy with the results, until Tango pointed out that R/O is not the proper way to reconcile Linear Weights.

BTW, the Linear Weights and Runs Created that I used for Chris’s equations were empirical. They weren’t derived from the partial derivative approach I just described.

Tango was correct in noticing that I didn’t treat the batting out (AB-H-K) component of the GIDP correctly. The batting out portion of the GIDP is now included at it’s full weight under AB-H-K. The advantage of doing it this way is that the formula is more context-neutral, since we are not penalizing hitter’s the full amount for their GIDP’s.

Rather than re-post the formulas, I posted it on Google Docs. The values in the spreadsheet get substituted into the Full XR equation. The dataset that the equation was derived from is 2001-2007.

http://spreadsheets.google.com/pub?key=pzy9IhjJPqatO2gC8b0DzOQ

So I checked.  It does use 0.72.  Well, that sucks.  So, I want it to be improved, so what’s the best formula.  Note, I know that baseruns won’t fall in.  Are terpsfans weights appropriate?  Also, I plan to use it for the last 20 years (1987-forward).  Should I use the weights in your Google Doc, or do you have the “combined 1993ff” weights handy?

Chris,

Can you post the exact equation you use, this way we know which parameters you are using, along with their coefficients.  (I am presuming you just want to alter the coefficients, while keeping the same parameters?)

Then, I, or someone here, can provide you with the best equation to use using 1987-present data.

Also, it would depend on how you feel SF, SH, and IBB should be handled.  Some people treat SH and IBB as “neutral” events, and give them either a league average runs per PA value, or a player-specific runs per PA value.

Some people think the SF should be treated just like any other out, while others think it should count with the knowledge that it scored a runner from 3B.

Once you answer these questions, someone here can prepare a DialXR equation.

I’m working on 2 XR equations for Chris from 1987 to 2007. The 1st is reconciled by R/O, the 2nd is reconciled by R/PA.

Here are those 2 equations:

XR RC1 = (.48 × 1B) + (.78 × 2B) + (1.06 × 3B) + (1.40 × HR) + (.33 × (HP+TBB−IBB)) + (.18 × IBB)+ (.18 × SB) + (−.28 × CS) + (−.082 × (AB − H − K)) + (−.114 x K)+ (−.30 × GIDP) + (.14 x SF) + (.10 × SH)

PAR RC1 = (Hitter OBP - LgOBP) * Hitter PA * Lg R/O

XR RC2 = (.60 × 1B) + (.90 × 2B) + (1.18 × 3B) + (1.52 × HR) + (.45 × (HP+TBB−IBB)) + (.30 × IBB)+ (.18 × SB) + (−.44 × CS) + (−.134 × (AB − H − K)) + (−.166 x K)+ (−.44 × GIDP) + (.06 x SF) + (.00 × SH)

RC1 = R/O Runs Created

PAR RC1 = Indirect Runs Created from the batter’s out-rate. If you add PAR to the a batter’s RC1 total, you can divide the runs created by plate appearances, instead of outs.

RC2 = R/PA Runs Created

The RPG from 1987 to 2007 was 4.68. The run-value of the GIDP is for the removal of the baserunner. The batting out of the GIDP is included under AB-H-K. I treated the batting out as the 1st out and the removal of the baserunner as the 2nd out. The .00 value for the SH in RC2 is not a typo.

Good discussion.  I wanted to make a point about seperating/combining the leagues for this kind of metric.

Metrics can measure two things - ability and value.  A pure ability metric would be a forward looking projection like PECOTA.  WPA is a pure value metric (although I would argue that in addition to the base state, out, inning and score variables they need to add a standings variable). 

Chris Dial’s stuff and VORP and whatever else all fall in a fuzzy middle ground.  Combining the leagues makes the metric lean towards the ability side, and seperating them makes them lean towards the value side.  I don’t think either choice is less legitimate than the other, but we should be aware of what the choice means

The precise formula for PAR is (Hitter NOA - Lg NOA) * PA * Lg R/O

where

Lg is abbreviation for League

NOA = (H + BB + HBP - CS - GIDP) / (AB + BB + HBP + SH + SF)

PA = AB + BB + HBP + SH + SF

Lg R/O = Lg Runs / (Lg AB - Lg H + Lg SH + Lg SF + Lg GIDP + Lg CS)

JBH,

It’s a good idea to have separate Linear Weights for each league. I don’t know if it has anything to do with ability/value. It is just more accurate to separate the leagues, because the run-values are different. Because the run-environment of the AL is always higher than the NL, Outs have a more negative value in the AL than they do in the NL. This also makes the run-values for hits and walks, a little bit higher in the AL compared to the NL.

terpsfan (and tango, et al),
thanks very much for your support.  I use outs used rather than PAs, so I will be using your first equation.  I’d like to properly credit you, so if you can drop me a line, that’d be terrific.

And thanks for the peer review, everyone.

Chris,

How do you plan on reconciling the XR estimates to runs scored for each league-season?

terps: if he uses the first one, how much will he be off each season?  At most 1 or 2%?  If so, then reconciliation for him might be overkill.

I mean, someone has 100 RC and it should be 101, compared to someone who has 80 and should be 80.8?  Works out in the wash.

I’m bothered by the K minus out gap.  Pretty substantial.

Can someone else confirm that?  Is it simply because of the GIDP (which I include in regular outs)?  I imagine that is the reason…

Tango - I calculated linear weights on 2005-2007
AL+NL data empiracally using even more out categories.  Here are my results.

K without SB or CS------(-.2912)
Air Out not DP-----------(-.2732)
Ground out not DP-------(-.2474)
Air out DP---------------(-.8999)
Ground Out DP-----------(-.8498)

So, yes there is that large a gap between hit ball outs and strike outs.  Terps values for DPs seem off to me though, if I understand correctly what he is doing.  It looks to me that he is adding in the +.12 value PA value twice, once when he calculates the batted out value and again when he calculates the DP value.  Or he may be doing some split of the DP value between batter and runner that I am not aware of.

Terpsfan is including the value of ROE/FC in the outs value. While, yes, those are considered batting outs the way he figures them (AB-H), it does lead to a distortion of the value of the out. I prefer to simply exclude those events and then adjust the reconciliation factor in order to make the events add up.

Terps may have included ROE and FC in the outs value, but I didn’t.  That isn’t the cause for the difference in generic outs value and strike out value.

This doesn’t take away from all the hard work that terps does, so please don’t take it personally.

I just want to say I’m not a fan of presenting the RC equation like terps is doing in part 2.

My preference is to show everything LWTS-style (runs relative to average), and then show an extra term to reconcile LWTS to RC, as .12*PA (or whatever it happens to be).  What we care about, what counts, is the runs relative to average (or relative to average per PA).

I think the presentation of merging the PA reconciler in one shot takes away more than it gives back.

Personaly, I would be in favor of dropping the term “RC” or “Runs Created” from the whole discussion.  It really belongs to Bill James and represents a flawed concept that nothing really needs to be reconciled to.  What we are doing by adding +.12 per PA is calculating Linear Weights above zero as opposed to Linear Weights above average.  The whole concept of trying to figure a player’s offensive value per out is just plain wrong.

What’s wrong with including the run-value of ROE and RFC’s under AB-H-K? That is what category they fall under. That is the reason for the .03 run gap between AB-H-K and a K. What am I distorting Colin? An AB-H-K is not always an out.

I had a discussion with Colin before, about how when I used the same categories that Tango used in the OUT category of his Full BsR equation, the value of the OUT was higher than the SO from 1994 onwards. So, I’m not calculating anything wrong here Peter. The value of the batting out for GIDP is included under AB-H-K. Had I included the batting out, under GIDP, then there would be a .04 run gap between batting-outs and strikeouts.

Tango: “I just want to say I’m not a fan of presenting the RC equation like terps is doing in part 2.”

Chris didn’t want an equation with marginal run-values, that is why I didn’t post it. Because, a static Linear Weights formula is of little practical use, why bother posting it.

You are correct that we should not be using the term “RC” or “runs created”.  You are correct that the concept itself, at the individual hitter level, is also wrong.

The wonderful think about runs above average is that it is terribly simple to convert to another scale (above zero, above replacement), simply by adding a constant to PA. 

That’s why simply showing runs above average, and number of PA are the only two dimensions that you need to worry about.  Everything else is derived from that (for individual hitters).

The way Chris does it, you don’t have to do any of that (assuming I’m remembering correctly). He takes XR for the player, and subtracts the XR of the average player at that position, given that amount of playing time. You could do LWTS Above Barry Bonds if you wanted to and it would all come out the same.

Patriot calculates RAA in the manner Colin just described.

Let me make a correction about the value of the OUT compared to that of the SO. The LW value of the OUT is higher from 1994 onwards, at least it is in the AL. Not the RC value.

AL LW OUT: -.309
AL LW SO: -.307
AL RC OUT: -.110
AL RC SO: -.120

NL LW OUT: -.286
NL LW SO: -.286
NL RC OUT: -.102
NL RC SO: -.113

Tom Ruane also included ROE’s and RFC’s under Batting Outs (AB-H-SO-GIDP). Take a look at the LW chart on Patriot’s website:

http://walksaber.blogspot.com/2008/06/run-estimation-stuff-pt-1.html

terps - Unless I am missing something about what you are doing your value for a batted out (-.134) + your extra value for a double play (-.44) should equal my value for a double play (-.85) + the value for an average PA (+.12).  Your -.58 doesn’t equal my -.73 and I can’t figure out why it doesn’t.

Peter,

Before I split the batting out, I had the GIDP valued at -.84 marginal runs. The only way they add up when you split the batting out and the baserunner out is if you assign the wrong value to the batting out. If you assign the correct run value to the GIDP batting out under AB-H-K, they will not add up. If you want I can try to explain it better or show you an example.

OK here is a detailed explanation about how I treated the GIDP in Chris’s XR equation. I am using the marginal values in the example.

Here is how my linear weights were initially categorized:

AB-H-SO-GIDP: 1723428 @ -.249

GIDP: 72969 @ -.842

Next I figured the value for the removal of the baserunner for the GIDP, which is the 2nd of the 2 outs.

GIDP Removal of Baserunner: 72969 @ -.438

I then figured the value of the GIDP batting out:

-.842 - .438 = -.404

The number of GIDP batting outs gets added to AB-H-K-GIDP, so our new category is AB-H-K.

AB-H-K-GIDP: 1723428 * -.249 = -429643 runs
GIDP Batting Out: 72969 * -.404 = -29479 runs

AB-H-K = 1723428 + 72969 = 1796397

LW Value of AB-H-K = (-429643 + -29479) / 1796397 = -.256

As you can see -.256 + -.404 do not add up to -.842, which is the total cost of the GIDP.

This method of treating the GIDP, does not penalize the batter for the full amount of his GIDP’s, which is a good thing, since we don’t know how many GIDP opportunities each player had. We are only penalizing the player for the removal of the baserunner.

terps - No, I understand what you are doing now.  You are basically double countiing DPs.  Once in your AB-H-K and again in a separate DP category.  It will add up to the same thing but it seems a crazy way to do things.  I agree with Tango.  Present your information as actual Linear Weights above average and then add the +.12 runs per PA to get linear weights above zero.

I have only one more thing to add about the GIDP. Because it’s only for the removal of the baserunner, I did not add .12 runs per PA to it, when I reconciled by R/PA.

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

Tango wrote this, and I agree with him except for one detail, which I describe below his statement:

“My preference is to show everything LWTS-style (runs relative to average), and then show an extra term to reconcile LWTS to RC, as .12*PA (or whatever it happens to be).  What we care about, what counts, is the runs relative to average (or relative to average per PA).”

I agree with presenting a separate term for reconciling with plate appearances (n*PA), but I do not agree with presenting a separate term for reconciling with R/O.

The R/O is not the same for all out categories. The R/O for the SH is approximately 25% higher than the average R/O. For the CS, the R/O is approximately 10% lower than the average R/O. So I would still prefer to present a separate set of Runs Created Linear Weights that are reconciled using R/O.