THE BOOK--Playing The Percentages In Baseball

It’s a shame it’s behind the pay wall...I said as much in the comments.  I’m glad my old boss let’s me piggyback on his account, I know I spend more time there than he does.

I get very annoyed with a few readers like:

“Great work, this is the direction fielding metrics should be moving in.”

The reader speaks as if he knows where the fielding metrics are. 

One good thing about the way Colin is rolling it out is that he’s bringing in the readers slowly but surely for the ride.  When MGL rolled out UZR for the masses, it was like he threw a tidal wave out there.

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

What kind of fielding metric is this? What data does it use?

I believe it uses the MLB Gameday batted ball location data.

To me, and please please correct me if I’m wrong, this seems like UZR except that rather than using just zones it uses angles.

So for every batted ball he is going to assign it an angle. Then based on the angle, he is going to compare how players do on ball hit on similar angles.

If that is the case, then it sounds pretty good to me as a start because the data would be more “precise”.

Angles would be more precise but Gameday’s hit location isn’t meant to be precise.

I don’t really recall Pinto’s methodology so much as his charts, which immediately sprung to mind when reading Colin’s intro.  What are the differences between this and PMR? Or is that what’s still to come?

RZ/6, I can’t speak to MLBAM’s intent, but there data is basically as precise as BIS or STATS batted ball data.  It does measure fielding location instead of landing location, but with regard to angles (as opposed to distances), that shouldn’t make much difference.

Dan/7, Colin mentions Shane Jensen’s SAFE system.  Since I can’t read the article, I don’t know where he went with that.  I can ask him, but I’m sure he’ll show up here to answer all this stuff himself soon enough.

there = their

Sometimes I hate not being able to edit my comments.

And I guess it was a commenter who mentioned SAFE.  I don’t know whether Colin referenced it in his article or not.

So they are all essentially just really good approximations of fielding location recorded by stringers and the location is not too far apart between BIS, Stats, and Gameday.

RZ/10 - yes, that’s a reasonable summary of the situation.  Peter Jensen had a more in-depth look into a comparison of the systems a couple years ago at THT:

http://www.hardballtimes.com/main/article/is-seeing-believing/

"To me, and please please correct me if I’m wrong, this seems like UZR except that rather than using just zones it uses angles.”

As far as I can tell, that is EXACTLY what it is.  I have said many times, that I wish that I had had the tools/knowledge to use each batted ball’s exact location and then to create a “smoothing curve” using LOESS regression or something like that. That is what Colin is doing.

On the other hand, a zone based system is to some extent a “poor man’s” LOESS regression, so I doubt that there is going to that much difference in the results, especially as the sample sizes get larger.

I would also like to point out, that using a “zone-based” system, like UZR, does not necessitate having “in-zone” and “out of zone” plays as ZR does and as Colin suggests that ALL zone-based systems do.  One has nothing to do with the other.  Some systems, like ZR, arbitrarily assign zones to fielders.  And some, like UZR, do not. Every zone is assigned to every fielder in UZR.  Some zones for some fielders obviously have zero catch rates.

Again, even though UZR is “zone based,” as far as I can tell, it does almost exactly what Colin is doing, except that it aggregates the data (into zones) and in addition uses the sample catch and batted balled ball values for each zone rather than constructing a LOESS regression-based smoothing function from the data.

Again, if you have large enough zones, and lots of years of data, a LOESS regression (to create that smoothing curve) becomes unnecessary.  Of course, by using large zones, you lose accuracy, especially in small samples of performance, as Colin illustrates quite eloquently and accurately with the green, red and orange dots.

Of course, as several people have pointed out, the less accurate the data is (in terms of the angle recorded by Gameday or whomever), the less you gain from using a LOESS regression function.  For example, if each data point were only accurate to within, say, 5 degrees one way or another, you essentially have a zone-based system anyway, and as I said, if your data is zone based, there really is no reason to do a LOESS regression.

Not to pry or anything, but does “SG” have a real name or another pseudonym?

Tango, that was me that annoyed you.  I was trying to be supportive of Colin but phrased it poorly and made factual claims without full understanding of the related fielding metrics and the math behind them.  In the future I’ll not make definitive statements about this area until I have a better comprehension.

MGL--

No, I don’t believe he uses anything besides SG (click my name).

jesse, cool, thanks for outing yourself!

Two ways to be supportive and not make any factually inaccurate claims:

- It’s fantastic that Colin is doing this great work in trying to bring BPro in-line with the rest of the leading fielding metrics out there

- Colin is giving us a great presentation in how to construct a fielding metric from the ground up

Right, there’s nothing particularly novel about what I’ve presented so far. (At some point I think there will be some things that are not, if not entirely novel then at least clever spins on the concept.) The element most directly inspired by UZR is the handling of plays made in split “zones” (although they really aren’t zones anymore). I don’t specifically mention that, but I do credit UZR as an inspiration and link to MGL’s primer on BBTF about UZR. Similarly, a lot of the other elements are directly inspired by SAFE, and that was credited as well.

PMR shares a lot of similarities with those systems, and so I can see the resemblance. (PMR has some ideas about park adjustments that I find intriguing, so you’ll probably see it mentioned explicitly down the road here.)

The concept of splitting ball responsibility the way I did I don’t know of a direct antecedent for. That’s really ancillary for now - UZR has no “rate stat” component that I know of, although MGL does discuss the idea of a UZR rate and a proposed method in the linked document that’s never been presented that I’ve seen. STATS ZR and BIS RZR both use a rate stat that differentiates between IZ and OOZ chances, which is what I was referring to when I talked about that.

Since I am using Gameday hit location data, I don’t have zones and so it made little sense to “force” the data into zones, for the reason I note. We’ll explore the question of how accurate the location data is along the line.

It is easy enough in UZR to construct a denominator which is the number of chances where each chance is simply equal to the percentage of balls fielded by each fielder in each zone (actually each bucket).  So if a certain kind of ball lands in zone A and the SS catches 60% and the 2B catches 20% (and 20% of the time, it is a hit), then the SS gets credit for 75% of a chance and the 2B gets 25% (3-1 ratio, like the 60%/20%). What you’ll find though, is that these chances are meaningless, although I guess you might be able to compare one fielder to another at the same position, but even then two fielders with the same “UZR percentage” or whatever you want to call it, will not necessarily have the same UZR in runs saved/cost, which is the only thing that counts.  So any kind of a rate (balls caught divided by chances) stat in a UZR-kind of system makes no sense since, again, every ball is technically a “chance” for every fielder.  I have written about this before.

Right, but you have expected outs for each fielder as well. So using odds ratio or log5, it’s relatively simple to take rate and “normalize” it so that you have comparable rates.

If you care to, you can even do it as runs instead, and then divide runs out by the average number of plays per run (or 1/runs per play, as it were) and then you even have plays normalized by run values.

I don’t know that it’s an idea worthy of much credit, but I find it useful (the Appendix to The Book gives one a lot of tools to use with binomials, and this gives me a binomial) and I don’t know that I’ve seen it done before, so I thought I’d incorporate it.

It sounds to me like MGL did it with his UZRrate, and I did it here among other places
http://tangotiger.net/UZR9903TT.html

Indeed, this is how I get my various error ranges.

That’s why I give a blanket 5 BIP per game for SS, 2B, 4 for Cf, 3B, and 3 for lf,rf,1b.  That’s how I get my opps.

I’m just happy that BP will finally be able to replace FRAA.  Every time one of the authors cites it, I get a little annoyed knowing that they could (and should) be doing better.

I’m pretty late here, but I just wanted to say that the primary benefit I’m seeing in Colin doing all of this is that we’ll ultimately have another data source to compare to the BIS-based data.  If we use the ZR conversions as well, then we’d have BIS, Gameday, STATS Inc, and Fan Scouting for any given player.  This would go a long way in helping us smooth out problems with the data sources.  I know others have used Gameday for fielding before Colin, but it’s not something that’s easily accessible for all players like bUZR currently is.  Looking forward to seeing this get put together.
-j

Justin, have you read Peter’s articles?

http://www.hardballtimes.com/main/article_preview/using-gameday-to-build-a-fielding-metric-part-1/
http://www.hardballtimes.com/main/article_preview/using-gameday-to-build-a-fielding-metric-part-2/
http://www.hardballtimes.com/main/article/using-gameday-to-build-a-fielding-metric-part-3/

He says he released 2008 data in the third installment.

Nick/23,

Yes, Peter is the main person I was referencing who has used gameday to build a fielding metric (and come to think of it, why didn’t we get that project nominated for a Saber??). 

My point was that Colin’s project, once complete, will presumably be implemented in BPro’s statistics pages to give us up to date data on players throughout future seasons.  Given the massive differences between sUZR and bUZR (and thus our awareness of how massively those two data sources differ), having another good-quality fielding metric available like this for all current players that uses a third data source will be very helpful.
-j

Yes, Peter is the main person I was referencing who has used gameday to build a fielding metric (and come to think of it, why didn’t we get that project nominated for a Saber??).

I did, but no one seconded. :(

http://www.beyondtheboxscore.com/2010/1/5/1234675/btb-sabermetric-writing-awards

Jinaz,

I did nominate Peter’s project under the novel research category.

Using Gameday to build a fielding metric - Peter Jensen

http://www.hardballtimes.com/main/article/using-gameday-to-build-a-fielding-metric-part-1/
http://www.hardballtimes.com/main/article/using-gameday-to-build-a-fielding-metric-part-2/
http://www.hardballtimes.com/main/article/using-gameday-to-build-a-fielding-metric-part-3/

I’m not sure which category this belongs under, actually. It’s novel research in the sense that it’s the first time anyone’s done this using a public source for batted ball location data, and that’s pretty important, IMO. On the other hand, the concepts of batted ball fielding metrics have been around for a few years in the form of UZR, etc. So maybe it should go under Applied Research. I’d be happy with it in either category.

Unfortunately, nobody seconded my nomination.

Mike/25,

Egg on my face, I guess.  I completely forgot that you nominated it, and in retrospect I probably should have seconded it since no one else did.  I think I ultimately didn’t because I was trying to avoid doing nominations or seconds in categories in which we already had large numbers of nominations.  At the time it was submitted, we already had 11 (though one was the cheeseburger thing and we opted to omit it, and two were for Allen and we made him pick one), and I was hoping to keep the number around 10 to keep things manageable. 

But it was deserving of a nomination.  Good work by Peter. 

Again, the main point is that the gameday work gives us a check on the BIS data, which is a really valuable thing to have.  If Colin’s work goes well and his metric seems to behave well--and given his track record, odds are good that it will--then I’ll be very happy to average his numbers with UZR (and maybe TZ and Fans) to get a nice composite fielding estimate for players.
-j

Thanks for the kind words and thanks Mike for the nomination.  I have been remiss by not getting the 2009 BZM fielding numbers published.  I have calculated them, and they will be published soon, I promise.  I have been working on other long term projects and I just haven’t taken the time to write the article to accompany the numbers. 

I don’t subscribe to BP so I haven’t been able to see what Colin has written so far.  I am glad that he is producing a metric using Gameday data.  When I published the articles about BZM last year I wanted to give enough information for others to use to also write a similar Gameday based fielding metric, but I purposely left out some of the details so that others could make their own choices as to approach without being biased by how I did things.  Fielding metrics are both simple and complicated.  Almost every fielding metric uses the same basic structure - how many plays does a fielder make given the opportunities that he has compared to others fileding the same position.  But many complicated small decisions have to be made that affect the final reported numbers for each fielder.  So far we have not had open source fielding metrics that use the same data that would give us a chance to discuss what these decisions are, and compare the effects of the different choices.  I hope the publication of Colin’s metric will give us that opportunity. I suggest that instead of just averaging the numbers as jinaz suggests in the previous post, that by fully discussing the differences in methodologies we may be able to make all fielding metrics better.  That, of course, depends on BP allowing Colin to disclose all the details of his metric.

I also have a Gameday based fielding metric to be published soon at THT.

I use some original concepts, was also influenced by Peter’s work, and has many similarities to what Colin has described so far.

There are some individual players where I disagree with UZR, and I will be interested to see how well I match up with Colin in trying to determine if the differences stem from the data source, the modeling, or both.

"I have calculated them, and they will be published soon, I promise.”

Awesome.

“I also have a Gameday based fielding metric to be published soon at THT.”

Awesome.

Brian - I thought that you did have a fielding metric from what you had posted in another thread, but I didn’t know what your plans were about making the information public.  Excellent.  Three different approachs using the same data should make for a very interesting discussion.  Can you reveal a tidbit in advance?  Is your system zone based or does it use a single large zone like BZM?