THE BOOK--Playing The Percentages In Baseball

Seeing as how your SD is half of what the SD of the actual sample is, I’d say that there’s a big chance that it’s not that your predictions have a poor correlation, but rather a sample size problem. If teams regress to the mean (which we’d expect), the SD would probably reduce as the season moves along, bringing your predictions closer.

What would help to figure that out, however, is if you would post a spreadsheet for each team so that we can see where the big gaps are between expected and actual performances by player to see who’s contributing to the massive gaps on teams like Toronto (34 difference), Houston (29 difference) or Texas (28 difference).

If, for instance, your projections are further off for newer players, then it’s possible that your projections did a poor job of evaluating younger/newer players. If the projections are off because a poor fielder got injured and replaced by a good fielder, or a good fielder by a poor one, then it’s more bad luck that may be resolved as the season moves along and players come back.

Any chance of sharing that data?

Hum. I’m still not sure if this means I did a good job (or a poor job) capturing team defense using the RZR data from THT:

http://otherfifteen.blogspot.com/2008/06/team-defense-in-nl-season-to-date.html

I’ll have to look over it again the morning.

Two questions on shortstops, if you’ll indulge me:

1) How does Ryan Theriot fare? Both from the data I have and from just watching him almost every time I see a game, I have the impression that he’s a very poor defensive shortstop.

2) J.J. Hardy of the Brewers - he seems to be the best shortstop in the NL if you look at RZR+OOZ plays, and the worst if you just look at STATS ZR. Is he good? Bad? Just in the middle?

Theriot was +36 (per 150) in 82 games in 07 and -3 in 50 games this year so far, for a total of +20 per 150 in 141 games, with no weighting by year.

Hardy was -1 (per 150) in 149 games in 07, +41 in only 26 games in 06 (+7 runs total), and -16 so far this year in 66 games (-7 runs total).

That is a total of -1 total runs in 241 games.

(I hate presenting “per 150” for a small sample of games, as the number looks big almost no matter what.)

I’ll see what I can do about putting together a spreadsheet, Sal.

Great.  Now you’ve given Trey Hillman another reason to keep Pena in the line-up. Like he needed a valid one anyhow.

It does not seem intuitive that a right fielder, no matter how good, could save 51 runs over 46 games. Right Fielders have, on average, made barely 2 plays per 9 innings this year with Gutierrez not making many more. How is Gutierrez saving those runs?

That is a rate of 51 runs per 150 games, as per MGL’s note.

If he’s played 46 games, then that’s +15 or +16 runs saved.

@JoeHova

I was also struck by disbelief at first by the individual numbers presented but after rereading the post I realised that the numbers for the best and worst fielders are (unlike the total numbers in the begining of the post) prorated to 150 games played.

Even so, a team of the seven best fielders (at each position) is more than 250 runs better than average which is significantly more than what I would have guessed.

Is that -20 in CF Andruw Jones or Adam Jones?

Jeez, Francoeur needs to be taken out back and shot, huh? He is just a sinkhole for the Braves this year. I’m disappointed but not surprised at Kelly Johnson’s poor score, and pleased with Yunel Escobar’s stellar ranking (I believe he scores highly in John Dewan’s system as well, which lends believability to the number).

tango-

unless I’m missing something, mgl specifically said: “Runs are total runs saved or cost, as of Saturday, June 7.  They are not “per 150 games.”

oh, ok, I missed the 2nd disclaimer. I see.  My mistake.

Guys, look at the top SLG at each position in any given month.  Is a team of those alot more than you’d expect?  Yes, of course.

That’s because the observed data is a combination of true talent and random variation (luck).  The fewer games, then the more luck impacts it.

Roughly speaking, you need twice as many games on fielding to get the same reliability as on hitting.  So, if you want to judge the range in UZR numbers after 50-60 games, look at OBP and SLG numbers after 25-30 games.  That’s the context you need to appreciate.

DK/8: marked for moderation and now unqueued.

Sorry, the numbers I presented are “per 150” and NOT total runs, as I said they would be.

That is why I hate giving “per 150” for small numbers of games.  For just a few runs saved or cost, the “per 150” seems like a large number.

Yes, that is Andruw Jones.  Unless there is something fundamentally wrong with UZR, and I don’t think there is, I am pretty convinced of two things:  One, A Jones has not been a good fielder for quite a while now, and two, Ichiro is not a good fielder in CF.  He was never had a great fielder in RF, according to UZR, and his UZR in center has been pretty bad.  This year it is -10 per 150 so far, last year it was -13, and in 06, it was +1 in CF and +2 in RF.  In 05, +1, 04 -1, 03 +14, 02, -7, and 01 +5.  05 and earlier were all in RF of course.

"Hardy was -1 (per 150) in 149 games in 07, +41 in only 26 games in 06 (+7 runs total), and -16 so far this year in 66 games (-7 runs total).”

I’m sure you’ve answered this too many times already but what is a “game”?  It isn’t games played or games started:

Hardy (GP/GS):
06: 32/29
07: 149/145
08: 53/52 (though 6/7)

Do you start with expected chances/game or something along those lines?

1. Figure out how many “expected outs” for all SS in MLB.
2. Figure out how many innings played divided by 9 in MLB
3. Figure out how many “expected outs” for Hardy
4. Divide #2 by #1 and multiply by #3

That’ll give you a “games-like” number.

MGL/14: I think Ichiro’s defensive reputation is based on his incredibly strong arm and his speed (and probably a little bit of his overall offensive ability because that’s how people think in baseball). People see a fast guy with a great arm and they assume he’s terrific in the outfield. Of course that’s flawed thinking, but that’s probably why.

People thought Vlad was a really good RF for a long time because of his arm, too.

Honestly, I thought that UZR simply underrated Ichiro, as some of the other defensive metrics had him as an excellent fielder.  I am starting to think that he is just not a great fielder, a little above average in RF and a little below average in CF.

I agree that an outfielder’s arm, although part of his defensive package of course (and Ichiro does have a good one), somehow gets “wrapped up” in his ability to catch the ball.  Same thing for speed.  Although speedy outfielders as a group are much better than slower ones (as you would expect), a fast and smooth, but otherwise not great outfielder can give the illusion of being better than he is.  As I said, I used to NOT think that was the case with Ichiro, but I am starting to change my mind.  Just too large a sample size.  Then again, SOMEONE has to be an outlier (his true talent is far from his sample performance), regardless of the size of the sample.  That is why Bell Curves have tails! Bottom line is that for all players, all performances, and all true talent levels, we only can make an estimate (which is actually a mean of a possible distribution) of a player’s true talent, based on metrics, scouting, physical traits and skills, or what have you. We NEVER know a player’s true talent until we go to heaven of course and God tells us everyone’s true talent, tells us whether clutch exists, etc.

Actually, now that I re-read the original post, there is no confusion about the “per 150” or not. I said that the team runs are absolute and not “per 150,” which they are.

Then I said that the player numbers are “per 150” which they are.

MGL, UZR and RZR/OOZ generally agree so far this year, with the Royals as the one big difference (26 UZR, -26 RZR/OOZ).  Is there something unique to the systems (or the Royals) that accounts for the difference?

I don’t know.  What is RZR/OOZ?  Is that just STATS ZR turned into runs saved/cost?

LF Duncan - 28?  This must be an outlier.  I don’t like to rely on obervation vs. good metrics, but Duncan has to be one of the worst LOOKING LF’ers I’ve seen in 30 years.

RZR/OOZ is the Revised Zone Rating/Out of Zone metric that Dewan developed and the Hardball Times currently uses. THT converts RZR/OOZ into a plus/minus scoring for each team.  KC (-26) was the one team that stood out in comparison to UZR.

How is that different from Dewan’s plus/minus?  His plus/minus is essentially the same as UZR if you turn it into runs.  The only significant difference is that it uses a different database, which seems to make a big difference.

Duncan was a -6 last year in 86 games, so between this year and last, he is +3 net runs in 119 games.  I am sure that the 95% confidence interval for 119 games would be like plus or minus 15 runs or so. 

As I always say, when you have a group of sample data, if there are not the usual outliers (whatever you would expect when you combine a bunch of bell curves), you are likely cheating.  Then again, we can’t just assume that every player who does not comport with our notion of that player is an outlier, otherwise, what is the point of the metric in the first place?  What we do is to regress the metric towards some “scouting” report on the player.  If that “scouting report” is a knowledgeable player watching a lot of that player, that is fine.  But, again, we have to be careful and understand that these metrics are generally better than our “minds and eyes” which can play tricks on us.  The metrics are (hopefully) unbiased.  We also have to understand that the larger the sample that the metric is based on, the more we have to rely on it, at the expense of what we “think we know” about that player (assuming that the metric has some requisite level of accuracy and reliability of course).

Duncan has some kind of a bad reputation as an OF’er, and I know he looks awkward out there, but LaRussa has said that he is actually a very good outfielder.  Whether Tony was just being kind or PC, I don’t know.  But when I think of Duncan, I always think of Tony’s remarks about his defense.

If I had to put a value on his defense, I would honestly say it was just a little below average (of course, with a lot of uncertainty) .  I get that estimate from his general rep (bad), his UZR (a little above average in 119 games), and LaRussa’s statements (which I interpret as being that he is average at worst).

Sometimes we also forget that a player’s defensive value is as compared to the average player at that position. Let’s say we take a player who looks good playing RF or LF and is a little above average (as compared to the average RF or LF of course).  Not let’s say that he moves to center.  Well, how he looks is not going to change!  He will still look good.  He will look as good as he did in RF or LF.  Yet somehow he becomes a below average defender!  How is that?  Most people, like commentators and the average fan, would think off the top of their heads, that he cannot go from being above to below average unless he plays center “worse” than he plays the corners.  That is not true of course.  He can play exactly as well in center as he plays the corners, and his defensive rating and value will go down by 5 or 6 runs (I forgot what the conversion is from corner to CF).  That is only because he is now compared against faster and better outfielders, not because he has gotten any better or worse.  People have a hard time with that concept.

It is the same thing as a player who goes from one league to the other.  In most of the last 10 years or so, the AL has been better overall in both pitching and hitting (although it looks like this year, the NL has become quite a bit better in the hitting department - the pitching in the AL is still a lot better).  When a player goes from one league to the other, he automatically gets better or worse (assuming there is a disparity in talent between the two leagues), only because a player’s value is as compared to the rest of the players in the league he plays for, at least for that year (of course, his “context-neutral” value is his value as compared to an average NL and AL player combined).

McAnulty, really?  He looks lost in left field.  Either mine eyes deceive me or the sample size isn’t big enough to capture his ineptitude.

This is what Cardinals fans thought of their fielders:
http://tangotiger.net/scouting/scoutResults2007_SLN.html

(Hint: go to the second to last line)

And here he is relative to all LF:
http://tangotiger.net/scouting/pos2007_LF.html

I will guess that if I take the top 10 Fans’ LF and the bottom 10 Fans’ LF and look at their UZR, then the top ones would be around +7 to +10 and the bottom 10 would be similarly around -5 to -10.

Has the Brewer’s infield defense been the worst in the league so far?  Both Fielder and Hall make the worst list and Hardy wasn’t far behind.  I’m surprised that Weeks didn’t make the worst list but it’s hard to believe he wasn’t close (he’s been a defensive liability since being brought up).

Tough to be a ground ball pitcher in Milwaukee these days.

MIL is -24 runs in the IF alone.  That is a lot.  That is .35 runs per game.  (The OF has actually been +9, so that should be a comfort for MIL fans.)

Branyan is +3 runs at 3B in 10 “games.”
Counsell is -2 runs in 22 odd games at various positions.
Dillon is -1 in 10 games at various positions.
Fielder is -6 runs in 58 games.
Hall is -6 runs in 46 games.
Hardy is -7 in 66 games.
Rivera is -1 run in 1 game.
Weeks is -4 in 69 games.

Yup, not such a good infield, especially the starters. I guess moving Braun to left field helped a lot, but they still seem to need more work on that infield defense.

Here are the same guys last year:

Fielder: -6 in 126 games.
Weeks: -20 in 105 games.
Hardy: -1 in 149 games.
Hall: -11 in CF in 122 games. (In 06, he was +7 in 118 games at SS, and in 05, he was +5 runs in 52 games at SS, so you would think he could play 3B just fine.)

Not a whole lot you can do with Fielder of course.  Weeks may need a change of position eventually.  Maybe CF, if he can play that. I don’t know.  I don’t know how long Cameron has left on his contract.

Despite the defense in the IF, the Brewers have a good team.  Losing Gallardo was a gigantic blow to them. He is a #1 or #2 starter.  Plus Bush pitching poorly and being in and out of the rotation did not help.  Plus I liked Capuano a little.  A decent #3 or #4 starter.  And Turnbow blowing up didn’t help.  And I would have kept Cordero. He is one of the best closers in baseball. I am generally not a fan of paying closers a huge amount of money, but if a closer is used properly, and averages more than 2 times leverage, and pitches at least 70 innings, he is worth a lot of money if he is a good one.  He is like a starter who pitches 150 innings. And Cordero is and was a good one.

THT converts RZR/OOZ into a plus/minus scoring for each team.

That’s not true.  We don’t convert Dewan’s RZR/OOZ data into anything.  Our team plus/minus figures are based on straight batted ball info (ground balls, line drives, etc.).  I would trust UZR and team-level RZR/OOZ data over the plus/minus figures we have on our team pages.

Just ran the correlation between my RZR/OOZ conversio to runs (NL-only) and MGL’s UZR numbers - .724. That’s better than I thought. I was way off on the Dodgers, though, it seems - I have them at plus 14. Yikes.

Does not mean you were off.  Dewan’s numbers are just as good as mine.  And there is lots of sample error in both metrics, and they will not necessarily overlap.

Apologies Studes.  There is no explanation of the team plus/minus on the glossary page, and the article linked on the glossary entry for RZR discusses the conversion of RZR/OOZ to approximate the plus/minus system, without clarifying those conversions are not the same as the plus/minus figures on the team pages.  But even looking at team RZR/OOZ, the Royals still rate pretty low (especially with OOZ), while UZR rates the Royals highly.  Other than the use of different data, is there something else going on here to account for the difference?

No apologies necessary.  The link to the explanation of the plus/minus stats is on the Team page (and that article links to the definition too).

Regarding the differences for KC, it looks as though there are differences in the source data.  MGL has Pena rated highly, but Dewan rates him 18th in his plus/minus system.  Teahen, whom MGL rates fourth, is 21st in Dewan’s system.

One thing I did this year, although overall it is only a minor change, is to eliminate all ground ball data when certain of about 12 batters are at the plate.  These batters usually have a shift on against them, and I think that with these guys at the plate, the data on who does or does not catch what in the infield is worthless, especially if you are using the usual baselines. 

The batters I am using now that I think most teams employ a shift on, are:

Giambi, Sheff, Ortiz, Delgado, Dunn, Howard, Thome, C Pena, Griffey, Hafner, and Fielder.

BTW, both Teahan and Pena both had a good UZR last year, Teahan +4, and Pena +9, per 150.

Any others that you guys know of?

I could not imagine that at least KC does not think that Pena is a good defender. With his “worst hitting in baseball,” in order for him to have any value whatsoever, he has to be considered a good defender.  UZR thinks that he is, and that makes more logical sense than he is not.  What did Dewan have for him last year?

MGL, in the BJ Goldmine, he has a list of hitters who get the shift the most.  It’s toward the second-half of the book.  For some reason, I have it in my head that Vlad was in that list.

OK, good, thanks.

When I get home tonight, I’ll see if I can go through the 2007 Gameday data I have and see who else might be seeing the wishbone shift regularly. My list of players who saw the shift was: Thome, Ortiz, Fielder, Giambi and Hafner. I’m curious to see if I can pick out Delgado, Pena and Griffey without telling the script to look for them specifically.

I think it should be easy for MGL to pick out the players, right?  In each zone, find out the position of the player who fields the ball, and see who stands out.  So, if the number of balls fielded at Retro zone 4 is 97% by the secondbasemen for the league in general, but is 60% with Ortiz at the plate (the other 40% being the SS or 3B), then we know that he was probably shifted quite a bit (like almost 40% of the time).

That’s what I do with the hit location data from Gameday, Tango. Of course, in that case, you wouldn’t have to eliminate those players from the sample - you’d just have to use a different set of numbers for each zone depending on whether or not you thought the shift was on.

Well, in some cases, it’s hard to tell, say if the CF picks up the ball where a CF would normally pick up the ball.  So, if you selectively remove some PA, you may be biasing your results.  That’s why I suspect MGL would simply remove all of them.

I’m not suggesting that any PAs are removed - simply treat it like any other special case.

For example - and these numbers are entirely made up - let’s say that in Zone G, 10% of the time the third baseman makes a play on a grounder. And when our population of shift batters is at the plate, the third baseman makes a play 60% of the time. You would simply calculate the player’s expected runs in that zone based upon the average chance of making an out based upon that population of batters, not the population as a whole.

Admittedly all I know of the workings of UZR is what mgl has published on Primer, and I’m sure that he’s made changes since then. So I’m fully prepared to be wrong here.

I don’t like the idea of using the whole population of batters who get a shift as a baseline.  For one thing, the samples are too small.  For another, each batter likely has a slightly different shift. I think you are going to generate a lot of noise if you try and include data where the infield is shifted in the UZR.

Tango, yup I guess I could figure out who is being shifted on.  Too much work though!  That’s a job for “game-watchers!”

I have also been meaning to remove situations where the infield was likely playing in with a runner on 3rd and 1 out.  I think I can identify them with 80% accuracy from the score and inning.  I might be able to use Colin’s idea and simply create a new baseline for the infield in.  After all, in UZR, remember I adjust everything for the outs/baserunners anyway.

p.297

69 Ortiz
66 Hafner
60 Bonds
56 Delgado
56 Thome
48 Howard
40 Dunn
37 Junior
34 Giambi

After that it drops down to 19 for Teixeira and 15 for Vlad.

The Nats employed the shift 44 times.  You’ll probably see alot of OOZ plays by Zimmerman.

MGL: you could use LI to figure out if the 3B is protecting the line.

MGL,

Any chance you have Matt Kemp’s UZR this year?

Question on Josh Hamilton (which is similar to the Hardy question at the beginning of the comments):

Right now, RZR+OOZ has him at…

90 PiZ - 94.1 ePiZ = -4.1
33 PooZ - 16.8 ePooZ = +16.2

+12.1 plays x .842 runs/play = +10.2

Would you say that the difference is due primarily to park factor, the shortcomings of RZR+OOZ, or some other factors entirely.

Thanks in advance.

The acronyms are getting out of hand.

No offense, but PooZ is a crappy stat.  Which reminds me, my baby is probably due for a diaper change.

Kemp in CF in 38 defensive games is -2 runs.  In RF, in 36 games, +3 runs.

Hamilton in 27 games in RF is +3 runs.  In 54 games in CF, he is -11 runs.

The difference between that and what you get from RZR+OOZ is most likely due to the distribution of plays and NOT park factors.  Park factors in any park but LF in Boston are not going to make a difference of more than a couple of runs or so.

Using straight STATS, Inc. ZR, I have Hamilton at -3 runs in CF and pretty much dead even in RF.

http://www.editgrid.com/user/cwyers/2008_ASB_WAR

IDs for the ZR data are STATS IDs; everything else is Baseball Reference.

(Use caution with the values for hitters; I used the NL replacement level across the board, because I was feeling lazy and I only needed Cubs players from the spreadsheet.)

I’ve taken a stab at actually projecting zone rating, using 2005-2008 data. There’s some regression to the mean, but no aging curve:

http://www.editgrid.com/user/cwyers/2008_ASB_ZR_Projections

I used 5/4/3/2 weights, and two weights for the league average. Plays/runs are figured over an entire season (not for the plays/chances listed), using Dial’s method.

If you have one list, and results are expressed in outs, can you then combine multiple positions?

Manny Ramirez is at the bottom - is he getting penalized for non-catchable balls off the Green Monster?

Certainly Manny is being penalized by the Green Monster, but I have no idea how to fix that; all I have are seasonal totals, and that makes it difficult at best to compute a park factor.

The problem with combining multiple positions is that outs aren’t equal between positions, either in value or in difficulty. I’m not sure how I would handle that right now.

Anyone know Orlando Hudson’s UZR?

It was a total of +1 runs in 75 games, as of a couple of weeks ago. Historically he is very good, around a +10 per 150 games +1 is right in line with that, of course, in only 75 games.

If you want a framework, having 80 games of fielding data would be like having 40 games of hitting data, or 10 starts from a pitcher.

So, if after a quarter of the season, Xavier Nady, as an illustration has the highest OBP+SLG in the league, that doesn’t mean that he’s now the best hitter in the league.  If a pitcher has a 2.50 ERA with a 6-2 record, that doesn’t mean that he’s now a great pitcher.

Whatever significance you want to give a hitter after a 40-game stretch, or a pitcher after 10 starts, that’s the significance you want to give a fielder after an 80-game stretch.

When you work with the UZR data like I do on a day in and day out basis and every year for 15 years or so, you easily get the idea of what kinds of fluctuations players have for no particular reason.

I am going to repeat something I have said several times already, but I think it is an important concept and it applies to all other sample stats, not just UZR.  It is more evident to me (the concept that is) because I work with the raw data.  It is not so evident to people who just see the results of the data (like a player’s sample UZR in x number of games).

Let’s say a player has a +10 UZR per 150 in the last 5 years, like a Hudson.  He is most likely a very good fielder.  We are not 100% sure he is, but we are 95% sure or whatever it is.  And of course, if we are going to attach a number to how sure we are, we have to specify a “talent interval,” such as, based on his 5-year sample of UZR stats, “We are 95% sure that he is a true +7 to +13 player.” Or whatever.  Anyway, that was a small digression.

Now let’s say that he is -5 in 100 games this year.  Does that mean that he, “Had a bad year?” IOW, does that mean that he happened to play poor defense this year (which is plausible, even if we are quite certain he is a very good fielder - just like a good hitter or pitcher can hit or pitch like crap for a while)?  No!

Or I should say, “Not necessarily.” Because of the nature of the data that go into the metric, there are 2 possible reasons for a good fielder having a sample UZR of -5 (or whatever “bad") in any time frame.  And we have no idea which one it is - it is most likely some combination, although I have no idea which one is more likely - I think it is the latter.

The two things are:

One, he did in fact play badly.  Let’s say that the balls in play near him were exactly the same distribution as they had been in the last 5 years, and roughly league average. But for whatever reason, or no reason at all, in this time frame (sample) he just didn’t get to the same percentage of balls he got to in the past or made more errors, or both.

Two, he actually “played” about as good as he has in the past and about at the same level as what we are pretty sure his true talent level is (very good), but the distribution of balls were “quirky” and made it look like he was a “worse” fielder than he actually was.  For example, say we have a bunch of balls in Zone 23 which was recorded as medium hit balls.  And we have our player fielding 40% of them whereas in the past he has fielded 50% (and say a league average player fields 45%).  Now, let’s say also that a lot of those balls - more than normal - in zone 23 were at the edge of the zone away from the fielder, so that in reality, NO fielder would field those balls more than 40% of the time.  Or say that a higher than normal proportion of those balls were actually “medium-hard” speed but were recorded as medium and he was supposed to, even as a good fielder, only field them 40% of the time.

Get it?  There are 2 sources of noise or what we call fluctuation in the sample data.  One is the actual performance of the player and the other is the limitation of the data.  We cannot distinguish between the two of them.  If we could, we could actually get a better estimate of a player’s fielding talent.  That is exactly what we do, for example, in FIP or DIPS ERA, or even when we look at batters’ line drive rates or BABIP (to some extent), or when we look at home runs by batters that just go over the wall or long fly balls that just miss being home runs.  Etc.

This same concept applies to batters and pitchers and anyone and everything that is being measured.  We have “measurement error” and we have sample performance fluctuation and at some level, depending on how coarse or fine our data is, they are indistinguishable, although we are constantly trying to reduce measurement error (there is nothing we can do to reduce the performance fluctuation).

This is a very important concept to understand.

For example, when Jeter’s UZR is +1 after 50 or even 150 games, that does NOT necessarily mean that he has “played better” than he has in the past or that he is having a “good” year.  It suggests that he is and has, but that is NOT necessarily the case.  Some of even all of that difference (between our prior and current estimate of his true talent and his sample performance) could be measurement error.  He could have played exactly as bad as he has over the last 5 years or so. 

So let’s stop saying that, “So-and-so is having a good (or bad) year.” How about, “So and so has X for a metric this year,” and then we can make anything we want of it.

Another example is simply when a batter hits a screaming line drive that is caught.  Is that him “playing badly” (it gets recorded as an out and you will never know that it is a hard line drive unless you are looking at that level of data) or is that “measurement error.” I submit it is the latter.  Same with UZR.  All you see if the final number (out or hit).  I try and make adjustments for the screaming line drive versus the bleeder through the infield, but I am limited by the data to some extent.

So when I see that Hudson is a +1 in 75 games yet I am pretty sure he is true +10 per 150, it wouldn’t take much measurement error for him to be a +1 in 75 games regardless of how he actually played.  In no way shape or form does that +1 mean that he has not played that good of a defense so far this year.

That is one of many reasons why when you work with this kind of data on a day by day basis, you pay little attention to what the numbers say on a short-term basis. When I look at UZR and I see some defender with a -10 or +10, or whatever, in 50 or even 150 games, I think to myself, “That’s nice.  Let’s see how he has done or will do in 400 or 500 games.”

Do you have Damon’s 2008 UZR in CF this year? How about Cano’s 2008 UZR? Thanks a bunch, love your work.

I am updating the UZR program right now in order to have better adjustments for the G/F ratios of the pitchers “in front of” each fielder.

I’ll have those for you later on today. IIRC, Damon had terrible numbers in CF for a small sample, and excellent ones in LF in a larger sample, suggesting that he is still a very good OF’er, probably CF-caliber.

Damon was +9 total runs in 82 games in LF and -6 in 30 games in CF.

Cano was -3 in 175 games.

Last year Cano was +9 and Damon was +6 (LF), +3 (CF).

In 06, Cano was +2, and Damon was +2 (CF).

So it looks like Damon is around an average CF’er, above average corner, and that Cano is a little above average at 2nd. At least based on the numbers.

thanks.

Not to start a string of individual player requests, but do you have Sizemore’s UZR (and if available, his actual plays vs. expected)?

Thanks.

Grady was +19 runs in 144 games and 368 chances.  A “chance” is the number of outs an average fielder would make. You can figure out expected and actual plays from these numbers. In 07, +17 in 167 games and 428 chances.  He is a monster player overall!  One of the best in baseball.

MGL - three questions…

1. have you considered a UZR confidence level per player?

2. has it gotten to the point yet where each OF grid has been absolutely defined so that you know the total sum of turf under foot each defender is definitively being measured on?

3. have you incorporated ball in flight times yet (flight times)? I’m thinking of cases where you have two teams with parks of relatively equal OF grid size but one team has to field 85% of their bip off flamethrowers vs another team that is fielding 70% of their bip off junk throwers.

#62 - Thanks!

One question - what do you mean by “167 games”?  Is it that he played so many innings that he effectively was on the field for more than a full season?

It means he had as many chances hit his direction as an average player would see in 167 games.

This may be slightly oversimplifying things but…

In 2007, Sizemore made 399 putouts and 2 errors for 401 actual balls gotten to.  If the average fielder given his distribution of balls in play would have 428 chances, how could Sizemore be +17?

Thanks.

MGL uses the word “chances”, which is confusing.  He is talking about “expected outs by an average player”.

"Chances” includes balls that could have been caught but fell for hits, not just putouts & errors.  Other outfielders must have a lot more chances - putouts than Sizemore did.

In post #62, MGL says:

“A “chance” is the number of outs an average fielder would make.”

Does that really include hits?  If so, how would you back into the “expected plays made” number?

If you have 100 balls in play where the average CF makes an out 90% of the time and
200 balls where the average CF makes an out 70% of the time and
500 balls where he makes an out 20% of the time
and 3000 balls where he makes an out 1% of the time
then… 100*.9 + 200*.7+500*.2+3000*.01 = 360 outs made on all balls in play

MGL calls this 360 number “chances”, when it’s clearly “average number of outs expected”.

For fielders, there’s no such thing as a “chance” equivalent to a PA, since each “chance” has a probability between 0% and 100%, unlike a PA where the chance of getting on base for the average hitter is 30% to 40%, depending on who is pitching and the park.

Right - I agree with that definition. So same question:

If the average fielder was expected to make 428 outs given the BIP distribution while Sizemore was in CF, and if Sizemore actually only made 401 plays, then how does he come out at +17?

I think MGL reported the wrong numbers. I have a previous version of his UZR, and in 2008 the average is 372 for Sizemore, and in 2007 it was 381.  It was 401 in 2006.

Birdo, you are right there is a discrepancy there.  I’ll check.  My numbers will be a little different than Tango has for 2 reasons.  One, I have been playing around with the adjustments for the ground/fly ratios of the pitchers in front of each fielder.  Two, I have been compiling UZR using BIS data.  I usually use STATS data.  It is shocking how different some of the UZR’s end up being.  There is only around a .720 correlation between UZR for the 2 sets of data across all positions (no one position correlates much better than another).  The numbers I reported here for Damon, Cano, and Sizemore are a simple average of the BIS UZR and STATS UZR.

Time in the air is not provided to me in the data.  I don’t know of any company that tracks that yet.  Maybe some of the teams are doing that themselves.  I don’t know.  Right now, the only thing that is tracked is the location of the batted ball (where it is touched or lands on the field), in terms of a point on the field (as estimated by someone in person and/or on video), a description of the “type”, where STATS only has grounder, liner, pop fly, and fly, and BIS has all those plus a line drive “fliner” and a fly ball “fliner,” and estimated speed - soft, medium, or hard.

As Tango says, there really is no “chance” for a fielder.  Every batted ball is a potential “chance.” So I merely define a chance as the number of outs an average fielder at that position would make given the exact distribution of batted balls when that fielder is on the field. From that, I compute “defensive games”. If a fielder’s number of chances is 300, and the league average is 3 “chances” (outs) per game at that position, then that fielder has 100 “games.” A chance or out always includes an ROE.

But Birdo is right that if Sizemore made 400 or so “outs”, his “chances” should be less than that.  I’ll check and report back.

What I like to do is take the “average number of outs” and divide by .7 to give me number of chances.

After all, DER = outs/BIP, and that comes out to around .70. 

So, if the average number of outs made is 350, then I count the number of “chances” at 500.

Instead of .70, you can use the league average ZR for that position, this way, you can compare UZR side-by-side to ZR. I think MGL did something similar several years ago, when he introduced UZR runs along with “UZR Rate”.

For Sizemore in 06, using the STATS data, I have 393 “chances” and using the BIS data, 418.

So it is in the ballpark for the number of outs and errors he actually had.  Tango is right in that UZR does not only look at balls caught versus how many would have been caught be an average fielder.  For example, if a ball is caught by someone else, no other fielder gets any demerits.  So, for example, if the fielders around Sizemore are catching a lot of balls, even if Sizemore does not actually catch as many as an average fielder, he could still have a plus UZR.

Got it.  So would the average DER be

(CF outs) / (All BIP - Other position outs)

or

(CF outs) / (All BIP)

?

Thanks.

MGL,

If you want to write those results up for THT, feel free to.

Birdo, let me think about that.  David, what results?

Birdo, I know what a DER for a team is, but I am not sure what you are asking.  The “DER” for a fielder could be outs/(outs expected by an average fielder given the BIP distribution for that fielder).  This would be a number near 1.0.  As we already discussed, this does not necessarily give you a great idea as to the proficiency of the fielder, as other adjacent fielders can influence this, as in the Sizemore example above.  I suppose you could adjust that number for the outs made by the adjacent fielders or the total outs made by all fielders in the vicinity divided by total BIP in that vicinity.

As far as the DER of the average, say, CF, I don’t know what that would be.  Outs divided by all BIP would be silly.  Even outs divided by all outfield BIP would also be silly.  As Tango mentions above, I used to do something whereby I gave a UZR “rate” in a number that looked like a ZR, for each fielder, as compared to an average fielder at that position.  I forgot how I did that though.  I’d have to think about it.  I think maybe I took each bucket and multiplied a fielder’s “ZR” (outs divided by BIP in that bucket) by the proportion of balls he fields in that bucket divided by the total number of balls fielded in that bucket, and then averaged all these numbers weighted by the total number of balls fielded by that fielder in that bucket.  Or something like that.

This is why using runs and the proportion of balls fielded by each fielder in each bucket is really the only good way to do it.  No other way really works.  One of the weaknesses with ZR, for example, among many, is that it arbitrarily assigns a “zone” to each fielder.  Even if you include OOZ numbers, you are still arbitrarily breaking the field down for no good reason.  (And of course ZR does not distinguish among balls in a zone or out of a zone - obviously within a fielder’s “zone” some balls are harder to field than others).  If you think that a metric like UZR is suspect (which to some extent it is, especially for small samples), then you HAVE to hate ZR, even if you include OOZ outs.  ZR is like BA (or maybe OBP) while UZR is like OPS.  One is so much better than the other, it is not even funny.

I may have slightly misspoke - I was not trying to get to a “Player DER”, I was trying to determine what the “proportion of balls fielded by each fielder in each bucket” is that you mention in the beginning of paragraph 3.

Is this fraction (CF outs) / (All BIP in bucket - Other position outs)?

I would think this would be the only way to keep a fielder from being dinged for another fielder fielding a ball. 

The only other way would be to calculate the proportion as (CF outs) / (All BIP in bucket) and then just exclude events with a field out from another player but I would think this would keep the aggregate +/- runs from equaling 0.

The results on the difference between calculating UZR with STATS and BIS data, Mickey. I’m fascinated.

David, oh, OK.  Yeah, as I said, I was shocked there were so many significant differences and that the correlation was so low.  I would have to look at the individual BIP’s to see what is going on. They are going to put the UZR’s from the BIS data on fangraphs soon.  If I wanted to write something up for THT, I’d have to ask David A. from fangraphs for permission.  He might like the publicity.  I’ll check with him.

Birdo, I don’t do any DER’s, so you can do it anyway you want.  If I had to do a DER for a player, I think it would be their outs made divided by the outs made by an average fielder given that fielder’s BIP distribution, but I’d have to think about it.

They are going to put the UZR’s from the BIS data on fangraphs soon.

Quite possibly the most exciting thing I’ve heard in months. (Sabermetrically, that is.)

Exciting indeed. David told me he wanted to do that at the MIT class, and I was waiting to see when he finally would.

It’s funny, while I am often the one to tout the PBP defensive metrics, if you look at some of the discrepancies between the UZR’s generated from the BIS and the STATS data, it becomes painfully obvious that no one should ever say anything definitive about a player’s defensive prowess from a UZR or other defensive PBP metric.  That is certainly true for one year or even two.  I have not looked at a comparison when combining several years.  The correlation should be better of course.  When people start making statements about defensive metrics based on some of these “poor man” UZR’s, that really irks me.  If you can’t “trust” a PBP one like UZR, how are you gonna trust some of these other ones?

Of course the same thing can be said about offensive metrics, even “pure” ones like lwts.  People think that these are so much better than the defensive metrics, but that is mostly an illusion brought on by the fact that we categorize offensive events as either hits or outs, and everyone treats those characterizations as the gospel.  What if we didn’t have hits and outs, or singles, doubles, triples and HR’s?  What if the only thing we had for offense was the batted ball info?  Ironically, that would be BETTER information, yet all of sudden everyone would be questioning the accuracy of offensive metrics like they do with the defensive (PBP) ones.

If someday, when a ball was put in play, every fielder got a “result” (like if the SS fields it, he gets an “out”, if he just misses it, he gets a “near out”, if he misses it by a little more, he gets a “not so near out” and if he had no chance, he gets a “hit"), all of a sudden everyone would love a defensive metric that used these “results” and their average run values (lwts for fielders), yet, this would be WORSE information then we have now!

Anyway, I am not saying that a PBP defensive metric is worthless or even very dubious.  I am merely saying that the differences between the results generated from two different data sources which “should” essentially be the same, highlights the amount of noise that is inherent in the data as well as the ensuing results of crunching that data.

I hope that David A did not want me to spill the beans about fangraphs and UZR.  He didn’t tell me not to, and I figure my mentioning it can only be good publicity for the site (not that I am doing it for that reason - I get no royalties, other than a nominal fee so that I can use the data).

MGL: the only thing that we want to see is an article on Ichiro.  I will guess that he’s one of the biggest outliers, and if so, then that’s where 100% of your article should focus on.