THE BOOK--Playing The Percentages In Baseball

The first thing I wanted to look at was this:
http://www.insidethebook.com/ee/index.php/site/comments/how_reliable_are_fans/#40

The Fans thought that the A’s, Mariners and Twins had the best fielding teams, while UZR had them as 9,11,16.  Pinto?  13, 17, 23!

The Fans thought that Cubs, Reds, Indians, Nats were the worst-fielding teams.  UZR had Cubs as the 3rd-best fielding team!  Pinto?  Cubs were 5th best fielding team!  HOWEVER, Pinto has the other three teams as the 3rd, 4th, and 5th worst-fielding teams in the league!

Hard to make much sense of it all.

***

At the team-level, the observed SD = .006.  Since we know:
obs^2 = true^2 + park^2 + luck^2
and since David has taken care of the park factor for us, and luck is 1 SD = .007
... uhhh....

Something is wrong here.  If the observed matches the luck, then we are saying that true talent has no impact at all on team fielding (or, that teams build their teams in such a way as to make sure they have offsetting players).

I’ve been doing this for a long time, and the luck portion was always smaller than the observed.

Right now, I think David made a mistake somewhere.  I’ll keep digging.

I’m converting David’s data into something more usable for me.  For example, given Melky Cabrera’s playing time, 1 SD = 7.7 outs.  He was +18.0 outs, making Cabrera’s z-score 2.32.  Manny was -3.46.  The standard deviation of all LF z-scores was 1.20.

Long-story short, the true talent distribution in LF is 1 sd = .021 outs per “ball in zone”.

Repeating for CF, 1 sd = .020 outs per BIZ.

For 1B, 1 sd = .038 outs per BIZ.  This seems rather incredulous, leaving us with Pujols at +39 outs and Ryan Howard at -27 outs, for a difference of 66 outs (about 50 runs).  This is a bit hard to accept.

Nonetheless, I’m excited to see more from David.

The link you posted doesn’t go through to David’s site. I think this is the one you are looking for ...

http://www.baseballmusings.com/archives/cat_probabilistic_model_of_range.php

Thanks, updated.

This seems rather incredulous, leaving us with Pujols at +39 outs and Ryan Howard at -27 outs, for a difference of 66 outs (about 50 runs).  This is a bit hard to accept.

***

Remember, David includes pop-ups and line drives in his model which allows for a lot of randomness and discretionary outs to sneak in.

That’s an interesting thought.  Let’s work it through then.

Let’s say our 1B had 400 BIZ, with 280 outs.  One SD = sqrt(.7*.3/400)=.023 per BIZ, or 9.2 outs per 400 BIZ.

Now, instead, let’s say of those 400 BIZ (280 outs) we have:
GB: 250 (180)
LD:  75 (30)
Pop: 75 (70)

And 1 SD is: 7.1, 4.2, 2.2 respectively.  If we add up the variances, and take the square root, we get 1 SD = 8.5, which is LOWER than the 9.2.

Why is that?  Because I have popups as almost sure outs.  By isolating as we are doing, we are reducing variance.

Therefore, the effect of treating all BIZ as having the same success rate is that it acts as a ceiling to random variance.

Someone can prove me wrong.

Tom,

You’re making one big mistake. Pop-ups aren’t 70 for first basemen. They’re 20-30. 70 of the pop-ups will be caught overall, but only 20 or 30 by an average first baseman. Except Howard might catch five, and Pujols might catch 60, because Pujols waves everyone off and Howard is always waved off. That’s why the number get screwed up.

Man, maybe Olney was right.  David’s numbers are jiving with Dewan’s nearly as well as I would have liked.

FYI, The Chronicler is looking at PMR vs. ZR:

http://anaheimangelsblog.blogspot.com/

David,

Are you saying then that the “expected outs” for a particular popout slice at 1B might be 0.95, and if Howard doesn’t make the play, but Utley does, then Howard get 0 actual outs with 0.95 expected outs , and Utley gets 1 actual out, with 0.05 expected outs?

That of course is silly, if true.

Tango:  that is what David is saying, and it is true of PMR (or has been in the past).  That’s why looking at GB figures only for infielders is critically important. 

My own view is that David P should credit a player only with the difference between the out and the total chance of a ball being caught by ALL players.  So in your example, Utley or Howard would get credit for .05, while the other player’s rating wouldn’t be affected by the play.

Correction:  in your example there would be no credit to give.  But let’s say the probabilities were:
1B out .60
2B out .25
RF out .10
Non-out .05

Pinto would give these net ratings if 2B makes play:
1B -.6
2B +.75
RF -.1

Whereas I’d give +.05 to whomever makes the play, and treat it as an opportunity only for the player who makes the play.  But if no out is recorded, then punish them all proportionately:
1B -.6
2B -.25
RF -.1

The other thing that annoys me to no end with Pinto’s presentation is that he insists on doing +/-DER, when no one in the world has any idea if +.005 is good or great or barely above average.

Since he shows the “exp” and “act” outs, why not simply show the differential as well?  Saying Pujols is +39 outs is alot clearer than anything else you can show.

Or, another way to show it is +14%, meaning that Pujols made 14% more outs than expected. 

Ryan Howard would be -27 outs and -9%.

***

I think I agree with Guy.

You got it, Tango. Guy is right.

I very much prefer Guy’s idea to how David Pinto does it.  But, like ZR and (I believe UZR), I think we’re better off neglecting infield pop-ups altogether.  There was a real issue with Orlando Hudson going nuts after pop-ups last season, and just going off the 1B data, it appears that he might have done the same thing this year (Dan Turkenkopf brings this up over at BTF).

Studes, when you say Man, maybe Olney was right.  David’s numbers are jiving with Dewan’s nearly as well as I would have liked., do you mean they aren’t jiving as well as you would have liked?

Oops.  Yup.

I also like Guy’s proposal for handling discretionary plays. In fact, he proposed this 9 months ago in a comment on David Pinto’s site. It’s unfortunate that David probably hasn’t taken his advice so far.

The other problem PMR suffers from at this point is tiny sample size. For outfielders, PMR has 60 different combinations of parameter value for each vector 5 hit types [pop, fly and 3 flavors of liner], 2 batter handednesses, 2 for for pitchers, 3 degrees of how hard hit the ball is. Park is an additional parameter, and within a given park there are only 60-120 balls hit to a particular (STATS) vector (except straghtaway center). [David gives a hint in his video presentation that he aggregates the BIS vectors to their corresponding STATS vector; reasonable since his model was originally based on STATS data; but if not the sample size problem becomes even worse.] With one year of data, for a given parameter combination, this year’s outfield PMR must usually be based on 5 or fewer outcomes for each parm combination. The situation isn’t as bad for infield grounders, but it would still be 10-20 outcomes for ground balls.

David G - if you meant fair popups only, there were an average of 31 per team by 1st basemen in 2005, but there also were an average of 46 more foul popups caught by each team’s first basemen. Admittedly there is much less “discretion” on foul pops. I don’t remember if it has ever been disclosed what PMR does with foul outs.

The other issue as well that needs to be thought through is year to year data consistency. If you look at the LD% it changed significantly between 2005 & 2005, I have no idea if this has happened in 2006. But where there is subjectivity about the metric we are always going to have some error in the results. (As Tango says we really need angle(s) of impact and hang time.)

***

On the sample size issue the data needs to be smoothed to make it work—I think MGL does this for UZR (either that or have larger zones). But it is a definite issue.

***

Guy, I completely agree with your handling of outs. Unless you do it that way I don’t think you can use the PMR model to compare the relative fielding win/run values for each player. I suppose if you get rid of pop-flys (which are guaranteed outs, largely) then this problem goes away to an extent.

John:  Pop-flys takes care of some of the problem but not all.  I imagine there are some soft GBs between 3B and SS that get fielded by both positions.  More importantly, there must be a LOT of FBs to the OF that could be handled by 2 of the 3 OFs.  A CF ballhog can end up looking a lot better than he really is.

Basically, my view is you don’t penalize any player for a ball that becomes an out.  In some cases that probably isn’t quite right—a great CF makes a tough play only because his weak LF couldn’t get there—but on balance you’ll get a much clearer view.

Agreed, Guy.

And if you have direction of hit, hang time and angle of the hit, then you can identify the “difficult fly balls / pop outs”.

May that time come soon ....

Pinto has SS up, and, finally, Betancourt looks great (#3, with a z-score of 2.2, and a +27 outs). 

I’d love to see how Dewan, MGL and Pinto differ on Betancourt, yet all agree on Everett.

***

Everett #1 with a z-score of 3.0 (+35 outs).  That is, Everett’s performance was 3 standard deviations from the mean.

In-between is Bill Hall:
http://www.tangotiger.net/scouting/scoutResults2006_MIL.html
who Brewers fans were high on (a bit above average for a SS), but they were even higher on JJ Hardy.  Will be interesting to see this next season.

At the bottom is Felipe Lopez, at -2.7 SD (-32 outs).  And the fans do have him as the 2nd worst SS in the league:
http://www.tangotiger.net/scouting/pos2006_SS.html
A shade better than Berroa, who Pinto has as the 7th worst.

Clint Barmes is another SS that the fans don’t like, but PMR and Dewan, both do.

***

The standard deviation of all z-scores is 1.25, which means the true talent is 1 SD = .016 outs per BIZ.

***

Except for 1B, the talent distribution is right in-line with what we think.

Given Melky Cabrera’s playing time, 1 SD = 7.7 outs.  He was +18.0 outs, making Cabrera’s z-score 2.32. 

How do you work out the 1 SD = 7.7 outs? David doesn’t give the BIZ data, only the number of outs and total balls in play when Melky was on the field.

Do you work out a weighted average DER for the total LF population and then apply Melky’s outs to it? Or do you somehow adjust for BIZ, in which case what is the BIZ for each player?

Thanks
Alex

A little trick of mine.  Take “expected outs” and divide by .700.  That gives you BIZ.  That is, the average ZR would be outs/BIZ = .700.

Once you have that, things go smoothly.

Otherwise, the way David does it, every BIP is a BIZ, from a ZR of .000 to 1.000. (That is, a ball hit to LF is a BIP to the 1B, with a ZR of .000).  Nothing wrong with that, but, now it’s pretty hard (though not impossible, if you construct a simple model) to calculate the standard deviation.  I prefer my quick way.

UZR loves Hardy (in limited # of games) and Barmes and likes Hall.  It does not like Berroa and hates F. Lopez.

And to change the subject and perhaps hijack the thread, I am not nearly as down on the Matthews and Pierre signings as many analysts as UZR has both of them rated highly on CF defense and with god CF defense usually comes good speed on the bases which is usually worth another 2-3 (or more) runs a year.

And speaking of Sarge Jr., where does this notion come from that 06 was this gigantic anomolous year for him?  The last 3 years lwts for Matthews was:

+14
0
+15

Granted in 03, he was -18, but in 02 he was +5.

It is funny that someone in the FO in Tex or maybe ANA said that they thought that Matthews finally “turned the corner” in his batting talent.  If that is true, he must have also “turned the corner” in 04, and then went back around the corner in 05 and then turned it again in 06.

I think that Sarge Jr. is a wel-above average CF overall and that Pierre is average or a little above overall.  Given the market for FA this year I don’t think their signings were bad.

I also laugh at the criticisms levied against signing players in their late 20’s or early 30’s to long-term contracts.  One, almost ALL FA’s are in their late 20’s or early 30’s and two, what you lose in production in age, you generally gain in salary inflation.  IOW, from an economic/performance perspective it really doesn’t matter the length of a contract, for batters that is.  For pitchers, I’m not so sure, although I don’t think they are all that much different than batters.  The main difference with pitchers is that they lose more TBF/PA per year with age than batters and of course they regress much more than do batters.

MGL, we have a lot of confusion over on the BTF thread on those Matthews lwts.  That’s not just hitting, that includes defense and baserunning and everything, right?  Because a +14 offensive lwts for Matthews in 2004 doesn’t seem to jive with his batting line.

It is offense only, park and opponent adjusted.  I’ll have to check the numbers when I get a chance, but I am 99% sure that it is correct.

In 2004 Matthews only had 280 at bats, with a .350 OBP and .461 SLG, which I think equates to about 0.4 above wins average over the season.

MGL typically shows things as “LWTS per 150 G”.  He may have done the same thing here.

David is continuing with RF and 2B.  He’s finally agreed to include distance travelled, though I’m not sure he’s doing it right.  This is what I posted on his blog:

There’s no question that distance travelled is an absolute requirement, and I would not believe that “hardness of hit” ball can even be close to a proxy. 

You also don’t want “distance/100”, unless you take the absolute value from the normal position.  That is, if a corner OF plays 275 feet from home plate, then you need to take the absolute difference from 275 feet.

I would also combine the distance ball travelled and zone, since all you care about is how many feet the fielder has to run, and how long does it take him to run that distance.

While keeping the same model for IF and OF may be nice, it’s hardly an overriding goal.

In short, scrap it, in favor of two models.

And, of course, you need to track XBH as well, since a guy who positions deep is purposely doing that tradeoff.

You *may* have similar issues with 3B.

The fact that the distance model changes ratings for 2B is a strong indication that distance has to be included, and also, as you say, that IF and OF should have separate models.  Grudz, for example, is ranked #21 by the original (velocity) model but #6 by the distance model.  That suggests the original model was holding him responsible for a lot of FBs to RF and right center—probably “soft” FBs that were hit far beyond the IF.  If velocity was even a half-decent proxy for distance, then the two models would give nearly identical ratings for IFs and only make a difference in the OF ratings.