THE BOOK--Playing The Percentages In Baseball

This is what you guys said pre-season:

http://www.insidethebook.com/ee/index.php/site/comments/poll_the_chance_that_strasburg_will_post_an_era_below_250_the_chance_he_wil/

And this was Lincecum:
http://www.insidethebook.com/ee/index.php/site/comments/poll_the_chance_that_tim_lincecum_will_post_an_era_below_250_chance_he_will/

***

I think it’s fair to say we’re going to see that the readers are going to be extremely bullish on Strasburg based on the one start, probably giving him a pre-season Lincecum forecast.

Oops, I misread the poll, I thought you were asking for a mean, and I voted 3.50.  That’s probably too conservative as a mean, so I would change my overall vote to 4.25.

The mean pessimistic forecast in March was 4.82.

The mean pessimistic forecast today (63 votes) is 4.18.

That puts his overall mean forecast (2.50 optimistic, 4.18 pessimistic) as 3.34, which is around 78% of league average.

That’s a pretty reasonable number.

After 101 votes, it’s 4.22 for the pessimistic forecast, and 3.36 for the mean forecast (78% of league average).

Mark Prior’s rookie year was 3.32 ERA (82% of league average), and Kerry Wood was 3.40 ERA (78% of league average).

It seems to me that fans have an implicit understanding of the uncertainty range of ERA.  Good job to them.

Couple that with ranking Strasburg in a similar pool as Josh Johnson and Verlander, and fans are able to put their virtual money where their mouth is.

Couple that with ranking Strasburg in a similar pool as Josh Johnson and Verlander, and fans are able to put their virtual money where their mouth is.

I bet if you ran that poll again, Verlander and Johnson wouldn’t be in Strasburg’s neighborhood.

I knew from PITCHf/x from the Arizona Fall League that Strasburg’s fastball and curveball were that good.  I didn’t know that he could keep it up for seven innings against major league hitters.  And I didn’t realize just how good his command was.  I didn’t realize how good his command of the changeup was.  Damn.

Are you seriously telling me that you think Verlander is that good?  Not a chance.

I’d have to think about Ubaldo with his great strides over the last year and his stuff, and whether Lincecum’s recent struggles mean anything, but I think anyone should have a hard time justifying someone other than Strasburg as the best pitcher in baseball.  Lincecum’s command is not where Strasburg’s was last night.

After what you’ve seen now, Tom, you’re still struggling to pull Strasburg back into the pack?

Roy Halladay?

David/6, certainly he’s in that discussion, too.  And Greinke, too, obviously, unless his recent struggles are indicative of something deeper.

Tango was referring to poll he ran about most valuable pitcher under age 26 (?).  So that’s why I/we were not discussing Halladay in relation to that. 

Before someone takes my comment far beyond what I meant it, I don’t mean that Strasburg is destined for a 1.00 ERA for his career or something.  My point was that on the basis of stuff and command, a pitcher’s tools, that you couldn’t really argue that anybody has better tools than Strasburg.

The young pitchers in the poll by the readers went like this:
Lincecum
Greinke
Felix
Jimenez
Josh Johnson
Justin Verlander

Strasburg, tools/body-wise, is the best of all the pitchers.  That is apparent.  Felix probably has the second-best set.  But, as we see with Felix, well, things don’t always work out.

The question on the table is which pitcher would you like to have for the next 4 years.

I put Strasburg in a pool with Johnson and Verlander, and I said probably behind Verlander.  Maybe after seeing this performance, I’d put him ahead of Verlander.  But, that’s pretty much where he has always been according to the fans, right, somewhere around Verlander?

So, I don’t know why you think I’m being pessimistic about Strasburg.  I’m pretty much like the average reader of this blog, which is controlled optimism.

***

Here’s the question: would the Giants trade Lincecum for Strasburg, today, if they got to hold either one for 4 years and 100MM?

Would the Royals trade Greinke for Strasburg, today?

Would the Mariners trade Felix for Strasburg today?

Jimenez?  Verlander?  Josh Johnson?

I think if you are fair, you would say that he’s in Verlander and Johnson’s camp.

Let’s see how he does the second time he comes through the league.  It’s possible that he will jump to ahead of Lincecum.  It’s also possible that he will drop back below Verlander/Johnson.

***

Here’s another one: just for the rest of THIS season, would you rather pay Strasburg 15MM from now to November, or Cliff Lee?

It reminds me of Doc Gooden in 1985… well, check out John Tudor in 1985 too.  From June 8 to the end of the season:
http://www.baseball-reference.com/players/gl.cgi?n1=tudorjo01&t=p&year=1985&share=3.52#150-174-sum:pitching_gamelogs

19-1, with a 1.32 ERA.  Speed and pitch variety helps.  But location is what matters most.

Speed and pitch variety helps.  But location is what matters most.

Right.  That’s what really impressed me last night.  I’d call it command rather than location.  I don’t think pitchers can really pitch to a location.  You can read what I said about Cliff Lee in the 2009 THT Annual to see my definition of command.

Strasburg displayed excellent command on all his pitches.  That’s what set Cliff Lee apart in 2008.  Strasburg has better stuff than Lee. 

The risk I see with Strasburg is (1) injury, but that’s the same for all pitchers, give or take, and (2) that his command isn’t really as spectacular as he displayed last night.  He will probably have nights where it’s not, but I don’t have any reason to think that the command he displayed last night isn’t a skill he possesses.

He’s also not going to strike out 14 batters every night, so he’ll have more balls in play, and some of them will fall in for hits or go out of the park just like they did against the Pirates.  That’s why any pitcher other than Sydd Finch can’t maintain a 1.00 ERA over a long stretch.

So, I don’t know why you think I’m being pessimistic about Strasburg.  I’m pretty much like the average reader of this blog, which is controlled optimism.

I didn’t say you’re being pessimistic about Strasburg.  Maybe that’s true, but it’s missing the point of what I’m saying.  I’m saying that you’ve been harping for a while that he shouldn’t be considered an outlier.  I’ve been harping all along that you shouldn’t be so sure he isn’t an outlier, when all the available evidence says he is, and after last night, my feelings on that are only reinforced.  It’s true that you’ve got history behind you on your side.  I’m not sure that’s enough. 

Right, it’s all a matter of turning the sample observation into true rates.

With fastball speed, you don’t need many observations to get to someone’s true fastball speed.

For movement, you need something more.

And for location (or command or whatever you want to call it), well, that’s the question: how many observations do you need to establish his true talent tool level for command?

(For example, what would you have thought of Felix at 19 years old?)

Add in pitch selection, and talent level with men on base, composure, and a couple of other things I guess, and we need ZERO performance based data. 

That is, you can forecast Strasburg purely on a scouting level.

***

My problem with the forecasters is that they focused mostly on his actual performance results (and so, did not regress enough, and did not build in the appropriate uncertainty level).

If you also include his toolset, then you have a better mean to regress against.

It all about population selection, regression, and uncertainty.

"I’m saying that you’ve been harping for a while that he shouldn’t be considered an outlier. “

He should not be considered an outlier based on the numbers, which is what all the forecasters have been using.

I also said that there’s a limit to what you can forecast someone (Tom Seaver) because of the unknowns.  What if his command isn’t as good as we saw last night?

It’s about uncertainty.  And location is the most important thing, not only being able to throw it to the spot, but also knowing which spot to throw to.

Strasburg looks, right now, as a combination of the good parts of Felix Hernandez and Cliff Lee, which basically would make him the perfect pitcher.  But, there has to be a certain uncertainty level that he’s got what Cliff Lee has, right, after one MLB game?

And that’s what I keep harping.  If you tell me he’s got say a 30% chance of posting a sub 2.50 ERA, well, then tell me what ERA level would give him a 30% chance of posting above that.

Uncertainty, uncertainty, uncertainty.

What I am harping about is looking at the optimistic AND pessimistic side, and saying that the true mean should be somewhere in-between.

Yesterday’s poll

Strasburg,2,6,4,2,80,Tangotiger,22
Strasburg,2,6,5,2,90,MarcoFujimoto,53

Marco was the median thru 98, so Tango’s pick was almost right at the median.

I was surprised the poll had so few K’s for Strasburg.

He should not be considered an outlier based on the numbers, which is what all the forecasters have been using.

Which numbers?  Do you mean his college and professional results: i.e., innings, walks, strikeouts, groundballs, etc.?

I specifically asked you whether we should restrict ourselves to only to those numbers/results, and you said no, that we should factor in all available information we believed was relevant.

If the restriction is to only use his results as inputs, then I agree that you are correct.

I also said that there’s a limit to what you can forecast someone (Tom Seaver) because of the unknowns.

But in other cases you’ve been asking us to forecast Stephen Strasburg only, and using all available information, not just that which is necessarily available for every player in a format that a forecasting engine could use.  In that case, I don’t see why you have to a priori impose a ceiling at Seaver/Gooden.

Strasburg looks, right now, as a combination of the good parts of Felix Hernandez and Cliff Lee, which basically would make him the perfect pitcher.  But, there has to be a certain uncertainty level that he’s got what Cliff Lee has, right, after one MLB game?

And that’s what I keep harping.  If you tell me he’s got say a 30% chance of posting a sub 2.50 ERA, well, then tell me what ERA level would give him a 30% chance of posting above that.

Uncertainty, uncertainty, uncertainty.

What I am harping about is looking at the optimistic AND pessimistic side, and saying that the true mean should be somewhere in-between.

Yes, yes.  Those are good points.  And I agree.  Though one small point - we have detailed data from three starts by Strasburg, not just one.

People were basically predicting (in yesterday’s context) a little more than a K per inning, based on the PC’s.  Tango had 6 K’s and 80 pitches.  80 pitches is like 5.1 innings. You can’t predict much more than that from a starting pitcher, no matter how good you think he is, can you?

Which numbers?  Do you mean his college and professional results: i.e., innings, walks, strikeouts, groundballs, etc.?

I specifically asked you whether we should restrict ourselves to only to those numbers/results, and you said no, that we should factor in all available information we believed was relevant.

When I say “forecasters”, I mean those people who spend the time to construct forecasting systems.  And those guys are not toolsy-based (or so as I have read it).

The readers were free to choose anything they wanted.

***

But in other cases you’ve been asking us to forecast Stephen Strasburg only, and using all available information, not just that which is necessarily available for every player in a format that a forecasting engine could use.  In that case, I don’t see why you have to a priori impose a ceiling at Seaver/Gooden.

Because of the uncertainty. 

If you are not going to forecast a confidence interval range for his true talent level, and you are only going to forecast a mean talent level, then you should not forecast anything better than Tom Seaver.

If you ARE going to forecast a confidence interval range, it better be larger than a young veteran’s interval range (Felix, Verlander, Lincecum, etc).

If you are going to tell me that you will forecast Strasburg with a better mean than Lincecum and a wider confidence interval range than Lincecum, then I want to see that forecast.  I want to see Strasburg’s 90 perecentile and 10 percentile forecast in print.

Until someone actually does that, then in no way can you forecast the mean of a player better than Seaver and at the same time with an implied wider confidence interval.

It’s the uncertainty level, the one that is implied/dismissed, that is at the heart of this issue.

The reason to not do it, the reason for the Tom Seaver rule, is to stop people with all their breathless forecasts of the annual once-in-a-generation player. 

How many players have been forecasted to be the best ever, and then actually become the best ever their first year? 

Imagine if we had the internet when ARod made his debut.  We don’t even have to imagine: Matt Wieters.

As you said, I have history on my side, which is the point: I’m using history to shape my expectations.

***

In hockey, it was Mario Lemieux.  He was exactly like Strasburg was.  A big imposing player who had all the tools to be the best scorer, best passer, and best skater ever (it was unusual at the time for such big sized players to also have all these skills).  And, he DID become all those things, he did become someone who was Gretzky’s equal, that he was arguably the best player ever. 

But, it did not happen in his first two years.  It sort of started his third year, but his coming out party was apparent for all to see in the 1987 Canada Cup.

But it was the same breathless talk.  I’ve seen it time and time again, and in hockey, players make a splash right away alot more than in baseball.

After 187 votes, and after we all witnessed that fantastic game, the pessimistic forecast is an ERA of 4.22 (against an optimistic of 2.50), for a mean of 3.36.

Only 3.7% of you think that his pessimistic forecast should be 3.00 or better.

I voted 3.50 for the pessimistic ERA, in case you want to gauge what I think from that.

I don’t disagree with what you see.  What I see is this: the pitcher with the best tools (stuff+command) of the PITCHf/x era.  Now, granted, stuff+command does not capture everything.  Moreover, the PITCHf/x era is very short: three years, going on four.

I just hesitate about where to put the cap on his upside given what I see from PITCHf/x.  Or to know where to realistically put his downside.  Obviously there’s a possibility of his downside being Francisco Liriano 2009, or something like that.  But how likely is that?

Frankly, I have a hard time having much certainty or confidence in your (or my) uncertainty forecasts, if that makes any sense.  They’re just guesses based on history, not anything about Strasburg himself.  That’s worth something, maybe worth a lot.

But then I look at the hard data we have from Strasburg himself, and I wonder how much the guesses from history apply.

You’re saying, I think, Tango, that on populations of “could-be-the-best-pitcher-we’ve-ever-seen”, the expectation is X.  And you’ll be right. And if I had to lay money, I’d lay it based on that population, as you’re suggesting we ought to do.  It’s balancing the risk against the reward, and you can’t do that very well when you focus on the data for the individual.

But that doesn’t make me happy as an analyst looking at Strasburg.  He doesn’t seem like he fits a population very well.

I wish we had PITCHf/x data for Seaver and Gooden and for Maddux and Clemens in their prime.  Then I could see if he fit in that population and maybe be a little more comfortable that the actuarial approach fits here.

If we’re going to project Strasburg on one start (yes, I know that’s not what we’re doing here but play along) then the one question we can ask is:

Of all the times we’ve seen a pitcher be THAT dominant how many belonged to Pedro in his prime?  I think we’ve seen Kerry Wood look like that on a few days.  I once even saw Sid Fernandez look that good (16K and 0 BB in 8 IP - you can argue that he didn’t have Strasburg’s stuff but on that day it looked pretty darn unhittable).  David Cone had days like that.  The VAST majority of pitchers never have a game like that.

Anyway, it was the best TV I’ve seen in a long time.

Mike, how do you define “command”?  If you’re looking at pitch f/x, you’re just looking at final location, right?

I’d think you would want to compare location to the location of the mitt.  For example, http://mlb.mlb.com/news/article.jsp?ymd=20100608&content_id=10944784&vkey=recap&fext=.jsp&c_id=mlb

Look at the first pitch.  Called a strike (which was a bad call anyways), but it was a solid 24 inches away from the target.  The third strike of that AB was about a foot low.  A great pitch, for sure, but I think he’d rather have put it where the catcher’s mitt was.

The pitch at 0:50 was supposed to be on the outside corner, but was about 6 inches inside.

The pitch at 1:19 was supposed to be a meatball, but missed by two feet.

Compare to Halladay: http://mlb.mlb.com/video/play.jsp?content_id=8495331

Granted, Halladay was “perfect” that day, but I can see only two pitch where he missed by 15 inches (in the 1st and 6th inning), and the vast majority are so close the catcher doesn’t really appear to move his glove.  Even on the contact plays, the catcher’s glove is stationary until the ball is struck, telling me the pitch was going to end up exactly where it was supposed to.  It’s uncanny, actually.

But compared to Strasburg, who on many pitches, was “effectively wild”, it seems like much better command from Halladay.

I watched all of the pitches in the second link from Mike above and I noted whether each pitch was pretty much to the intended location, within a few inches (I), missed by a little, probably 3-12 inches (L), or missed by a lot, probably more than one foot (A). I also gave more leeway to the curve ball.  Here is what I noted:

1) A
2) L
3) I
4) I
5) I
6) I
7) L
8) I
9) A
10) L
11) I
12) I
13) A
14) I
15) L
16) I
17) L
18) L
19) I
20) I
21) I

So out of 20 pitches, he hit his spot pretty much on 12 of them, or 57%, was off by less than one foot on 6 of them, or 29%, and off a lot (more than a foot, probably close to 2 feet), on 3 of them, 14%.

I have no idea how that compares to an average pitcher, or a pitcher like Halliday or Lee (or Maddux).

I have watched thousands of games in my lifetime.  If I had to say what kind of command Strasburg had after watching him pitch last night’s game, I would probably say, “Pretty to very good but not great.”

I have Dennis MArtinez’s perfect game on disk, and I’ve been meaning to do a detailed pitch-by-pitch, play-by-play… you know, the ideal sabermetric scorecard.

I had forgotten about including the catcher’s target.

I’m going to try an inning, and see what I can do.

My sense, from the games I have charted, is that missing the catcher’s target by 1-2 feet on any given pitch is average.

That’s from a small sample, say three or four games, and iirc, it was Brian Bannister and Cliff Lee, who are better than average in terms of BB/9. 

It appeared to me that Strasburg was doing better than that, and MGL’s charting of those 21 pitches reinforces my opinion.

Our understanding of command is really in its infancy in sabermetrics.  So when I say I think Strasburg is among the best in MLB right now, that has a huge error band around it.  Probably wide enough that what I am really saying is that if we divided MLB into thirds, Strasburg would be in the best third.  I don’t know that I could distinguish any pitcher, other than Mariano Rivera, who I could say would be better than anyone else in the top third.  So I consider that a high compliment to Strasburg.

If someone were to chart catcher targets for a lot of games, say 20 or 30, they would probably have a better feel for this subject than me.  Get back to me in a year once you’re done with that.

My expectation for Strasburg prior to his debut was that he would be in the middle third of the league in command.  That’s based on the PITCHf/x data we had from two of his starts in the AFL plus the reports from scouts that he had good command plus his low walk rate in college and the minors.  A typical hard-throwing minor league prospect I might expect to be in the bottom third of the league in command in his debut start.

So I was pleasantly surprised to see, as MGL called it, “pretty to very good” command.  Putting 57% within a few inches I consider very good.

It may be the Halladay’s command is even better.  I haven’t charted any of his games.  Like I say this is not a subject that is explored very much by sabermetrics.

I’ve talked qualitatively about this so far, but I suppose I should dig into my archives and find some of my quantitative data.  For Brian Bannister’s July 9, 2008 start, I charted 97 of his pitches.  His average miss relative to the catcher’s target was 1.1 feet.  Here are the results:

4 within 3 inches (4%)
44 within 3-12 inches (45%)
27 within 12-18 inches (28%)
13 within 18-24 inches (13%)
9 within 24-36 inches (9%)

So to Strasburg’s 57-29-14, Banny had a 4-45-51.  That may not be exactly a fair comparison since there were two different classifiers (MGL and me) in operation, but it does put some numbers to my claims.

Btw, if you expand the toughest bin to be within 6 inches, Bannister put 20 of 97 pitches within 6 inches of the target.

Also, I realized this might come across differently than I intended it:

If someone were to chart catcher targets for a lot of games, say 20 or 30, they would probably have a better feel for this subject than me.  Get back to me in a year once you’re done with that.

Anyone can certainly have a valid opinion and good ideas.  I’m not trying to say I’m the only one who knows anything about this subject. 

What I was trying to say was twofold, and it probably suffered from combining the ideas into one sentence.  (1) People have a much more optimistic idea of the ability of a pitcher to spot a pitch than the actual facts will reflect.  Charting a few games will quickly disabuse you of this notion.  (2) If someone wanted to get a quantitatively valid idea of the distribution of command among major league pitchers, they would need to chart a lot more games than I did, maybe 20 or 30.  That would take a long time and a lot of effort.

I would definitely not compare MGL’s numbers to Mike’s numbers.

In my limited viewing experience, Strasburg’s movement is more impressive than his command.  Some umpires would have squeezed him on that fastball that was tailing away from the lefthanded batters, pushing him into deeper counts.  Rookie pitchers simply do not get that call.  Maddux and Pedro got those calls, but they had already established themselves as the best in the game.  It will be interesting to see how Strasburg responds when he gets a tighter strike zone called.

I’m envisioning a high school research project for next year involving measuring velocity and accuracy of the team’s pitchers (throwing at targets on carbon copy paper taped against the wall) and then tracking game results throughout the season.  It won’t answer the questions posed here but could be fun.

As we move further away from results-oriented analysis (like ERA), we realize that our metrics can only take us so far.  Ideally we would like to evaluate a pitcher on his command (as it is defined in this thread).  At least...I would.  I would like to know instead of the result of a pitcher’s performance, how he was able to command/control his pitches as he had intended to, and furthermore I would like to assign various probabilities to the outcomes.  If a pitcher makes a mistake, it is luck that determines what happens to the pitch.  Of course.  That holds true in any count and in any situation.  Say a curveball is supposed to be down and away in a given 1-0 count, and it hangs over the plate.  Maybe 25% of the time it is taken for a strike.  10% of the time it is swung-on-and-missed.  15% of the time it is fouled off.  And then say 10% it is a HR, and the rest it is in play.  Well, when a pitcher gives up a HR, he of course was unlucky, because even though he made technically made a mistake, the expected value of his mistake is certainly not +1 run, since a HR happens only 10% of the time.  It’s why even if a pitcher throws 4 fastballs down the middle for homeruns, and prior to throwing every pitch declares “it’s coming right down broadway,” he has been very unlucky to surrender four homeruns.  Even if a batter has an extraordinary 50% chance to hit a HR in that spot, our poor pitcher’s still been unfortunate.

But that type of analysis, as Mike says, would take a tremendous amount of time and effort, and the payoff cannot be worth it.  We have many good statistics that may not have this level of granularity but which perform the same function (like FIP).  And, besides, the assumption lurking behind statistical analysis is that, over time, these things will “even out.”

"I would definitely not compare MGL’s numbers to Mike’s numbers.”

Definitely.  For one thing, my definition of “within a few inches of the intended location” is very loose for the curveballs and changeups. 

For example, say the count is 0-2 and the pitcher throws a curve or change up.  For one thing, there really is not an exact intended location.  For another, the pitcher often wants to throw that pitch in the dirt, but the catcher is not going to put his glove on the ground.  So if I say that the pitcher hit his intended location on a curveball low and away, it could be in the dirt or it could be 6 inches off the ground.  Either one is a good pitch which went where it was supposed to.  If it was a little toward the middle, it would be a small miss.  If it was a little high it would be a small miss.  If it was inside or right in the middle it would be a large miss or if it was thigh high or higher it would be a large miss.  I suppose if it bounced on the plate or in front of the plate it could be a small or large miss, but that would depend on the count and the batter.

In fact, the only pitch where the intended location is right where the catcher’s glove is is the fastball.  And even then, for a sinkerballer, there is probably a 6 inch “window” which indicates perfect location.  And when a pitcher throws a high fastball, the ball could be shoulder high and be a good pitch (intended location) or it could be letter high and be a good pitch (intended location).  With an 0-2 count there really is not an exact location for a high fastball.

So charting pitches and noting whether they were right there, a little off, or a lot off, is very subjective and without specific rules, can vary a lot from person to person.

So I definitely would not conclude anything from Mike’s and my numbers on Strasburg and Bannister.

Pitcher command to me is a little like obscenity to SC Justice Potter Stewart.  “I know it when I see it.”

"In fact, the only pitch where the intended location is right where the catcher’s glove is is the fastball. “

Right. Maybe.

I would say the best way to prove this is to take those guys who get great results from their various pitches (say a Lincecum changeup, Wainwright curve, Jimenez fastball, etc) and see where their catchers set up and where their pitches go.  If there is little correlation, for example, in Wainwright’s curve and where the catcher is setup, but the batter if flailing away or the umpire calls strike after strike, well, that means that the catcher setup spot is only used as a frame of reference for wainwright, and not a final target spot.

After 296 votes, we’re very stable:

pessimistic ERA = 4.22

midpoint ERA = 3.36

Poll will close in a day.  I’ll run a new poll following each of his starts.

I have great admiration for anyone willing to invest the time in charting enough games to try to measure intended locations.  But it seems like an area where even a lot of data is likely to yield inconclusive results.

An alternative approach would be to evaluate pitchers against the average run value of the locations they throw to.  (I haven’t read all of Mike F’s stuff, so apologies if this is what he already does.) We know how well hitters perform on each pitch type in various buckets of the strike zone.  So measure command in terms of the average run value of the locations they in fact pitch to.  In fact, we don’t really care about intent at all—if the pitch goes to locations where hitters can’t hit that pitch type, he’s pitching well.  So why spend a lot of energy trying to measure intent (which may or may not be possible)?

And once you have some data on a pitcher, couldn’t you infer that location is responsible for whatever you can’t account for with velocity and movement?  That is, a 92 MPH fastball with horizontal movement X and vertical movement Y should have an average run value of Z.  To the extent that a given pitcher has outcomes on that pitch that are better or worse than Z (controlling for count, and hitters if necessary), then can’t we say his location accounts for that difference?  Again, I don’t really care at that point whether he’s throwing it where he “intends” to—if it’s more effective than average, then he’s throwing it to good locations.

Guy, fascinating thought!

We only care about intended location if we are trying to give advice to the pitcher.  If we don’t have any advice to give him, then the only thing that matters is where the pitch actually ended up, compared to where the batter expected the pitch to end up.

Love the idea of trying to back out the information.

Guy’s idea sounds alot like defense metrics - create zones, calculate the expected value in each zone, and compare that to the observed value. The wheels are spinning inside my mind

The only problem with that is that we are trying to use intent and actual location to separate out the luck - so that we can infer true talent from small samples. With Guy’s method, you have not separated out the luck.  As an extreme example, if the wildest pitcher in the world (with horrible true talent) happens to have thrown 3 great pitches in 3 perfect locations by sheer luck, we still have no idea of his true talent.  We think he is a great pitcher when he is not.  Aren’t we trying to separate luck from talent?  What is the point of using actual location?  Sure, in the long run, that is all we need to know, but in the long run, all we need is ERA too!

Also, the difference between actual value and expected value by speed and movement is NOT just location.  It is location plus deception plus the sequence of pitches plus the mixture of pitches plus error in the measurement (of speed and movement) plus luck.

Bottom line is that we don’t need pitch f/x to infer true talent from a (really) large sample of data.  What we are trying to do with pitch f/x is to infer true talent from smaller samples of data. One way to do that is to figure out actual location versus intended location (there is obviously luck in that too, so we still want large samples of data).  There is no substitute for knowing the intended location.  None.

All these methods by Guy and Mike Fast are simply other methods of inferring true talent. They are not proxies for intended location.  There are no proxies for that.  Using Guy’s and Mike’s methods, you might be able to infer X true talent with a standard error of .1 rpg (that number made up of course).  If you knew the intended location for the same sample of pitches, you would have an inferred true talent of Y, where Y is more accurate than X, and your standard error would be less than .1.

"If you knew the intended location for the same sample of pitches, you would have an inferred true talent of Y, where Y is more accurate than X, and your standard error would be less than .1.”

This is much more wrong than right.  We wouldn’t “know” the intended location—we’d have an estimate of it.  And for a vastly smaller sample of pitches than we have for actual location.  And of course you would still have luck mixed in with the talent, just as you have when you look at location. 

If a curve from LHP to LHH is most effective in location X, and a pitcher manages to throw his curve to X more often than most LHPs, then I’m happy to conclude he has good location.  If he did that while “missing” the catcher’s mitt, it could mean he got lucky.  But it could also mean you measured it wrong, or the mitt isn’t a good proxy for intended location.  Really, who cares?  It would take years, at best, to accumulate enough data to have any idea.

We don’t know if a hitter had a “good approach” on any PA. Maybe he hit a double, but was just guessing and got lucky.  But we don’t care—if he hits enough doubles, we decide he’s a good hitter.  If a pitcher throws to objectively good locations, I don’t care whether he intended to or not.  And I bet measuring his intent will improve your ability to predict the quality of his future locations barely at all. 

A lot of work has already been done to map run values for each pitch type by location.  It’s a relatively easy step to compare a pitcher’s performance against that.

I agree that on the second method, the residual run value after velocity and movement would include pitch sequence as well as location.  One could try to control for that, but that would require quite a bit of data.  If you rated location using the first method, you could then see if there was any skill left after accounting for speed/movement/location (which would mainly be sequencing). 

“There is no substitute for knowing the intended location.  None.”

This is actually similar to the PBP fielding issue:  the assumption is that more micro-level data must, MUST, be better.  But given the potential for measurement error, it often simply isn’t true.  In this case, I bet looking at location will give you 95% as much information, at 5% of the cost.

"Guy’s idea sounds alot like defense metrics - create zones, calculate the expected value in each zone, and compare that to the observed value.”

Brian:  You’d calculate expected values in zones, just like in fielding.  But then it would be analogous to rating hitters based on how successful they were in hitting the ball to hard-to-field locations ("HZR").  In this case, you’d be rating pitchers on their success at pitching to hard-to-hit locations—as determined by results for many pitchers, not just themselves. 

*

Another problem with measuring intent is: what do you make of pitchers who “hit their target” but throw to poor quality locations?  I guess we’ll then create a separate metric for “target selection.” And then we’ll have a debate about how to allocate responsibility for target selection among the pitcher, the catcher, the manager, and the pitching coach.  But will the data really ever be good enough to distinguish between pitchers who try to throw to the correct location but miss, and those who have good command but are stupid?  Compared to this, determining true-talent platoon splits for players is a piece of cake.  Better to just try to figure out if they’re throwing it where they should....

In this case, you’d be rating pitchers on their success at pitching to hard-to-hit locations—as determined by results for many pitchers, not just themselves.

I think you’d want to find the hard-to-hit locations-as determined by the results for pitchers with similar velocity (and possibly movement).

An inside fastball to a RHB may be bad location for a soft-tosser but good location for a hard thrower (I believe I read this somewhere but even if I’m messing up the details you get the idea).  Of course then it becomes harder to compare apples to apples in your command metric.

I really like Guy’s ideas here.  What I see is the need for a 5 dimensionally smoothed equation for pitch value: vertical location, horizontal location, speed, horizontal break, vertical break.  With enough data, such a smoothed pitch value estimate could be obtained, couldn’t it?  I’m not sure what type of regression could do it, though.  Could LOESS be used to generate large numbers of “heat maps” for multitudes of pitch types?

It turns out Jeremy Greenhouse has already done this for RHP’s fastballs:  http://baseballanalysts.com/archives/2010/06/a_shot_at_comma.php.  Results seem consistent with expectations.  (I swear I hadn’t read it when I made this suggestion.)

Whether this approach is better/worse than a clustering analysis, I don’t know.  To some extent, a pitcher should throw to a variety of locations.  So a very tight cluster could be self-defeating at some point.  Of course, for the same reason a pitcher should sometimes throw to “bad” locations.  Once we have more data, we should be able to validate these metrics by seeing which better predicts a pitcher’s run values, once we’ve controlled for velocity and movement (assuming that sequencing is an independent skill).