THE BOOK--Playing The Percentages In Baseball

His 10th percentile PECOTA forecast has him hitting about as well as Derrek Lee/Raul Ibanez/Torri Hunter did last season. Wow.

That of course is complete b-llsh!t.

My number 1 problem with the PECOTA percentiles is that it presumes the range in forecasts is NOT tied-in to the uncertainty level of the mean forecast.

I did this when they came out years ago, and you can try to do it yourself: pick out all the top late-20s players in baseball (pitching and hitting).

Then, find the top rookies they have forecast.

Look at the 10%/90% percentile rankings of both groups.  You will see that they are very close in terms of breadth.

And that’s ridiculous.  Won’t you know more about the mean of a player’s forecast if it’s Utley, Sizemore, Mauer, etc, than if it was Wieters, and any rookie this year?

I haven’t checked if Nate changed it this year, but I presume that he hasn’t.

***

The reason it works like this is because Nate looks for the top comps first, and THEN be builds his uncertainty.  HOWEVER, the uncertainty has to be based on the sample size of the performance of the player and not on the sample size of the performance of his comps (well, that too, but the former far more). 

It’s an egregious error as far as I’m concerned.  No one talks about it because Nate has never come out to claim anything great about his percentile rankings.  And readers are so vowed by the very idea that a forecast could have an uncertainty that they blindly accept that the uncertainty must be based on logic.

Like I said: b-llsh!t.

I’d rather that BPro would remove the percentile rankings until they actually show evidence that they work.  Nate said they tested it a couple of times, but his test did not test for the bias I am talking about here.

PECOTA’s percentile rankings do not work, and I look forward to BPro to show me the evidence that they do work, with specific regards to the career PA bias I am talking about.

I won’t claim that mine “work”.  I’m not even sure what that means, but the % projections on my site do take into account the size of the sample the projection is based upon.  You will see a tighter range of forecasts for MLB veterans (Vlad, Torii, BobAbuee) than for the Angel rookies.

Well, at least in batting and OBP.  The power hitters will have more absolute room in their slugging forecasts, though not on a percentage basis - Vlad may hit anywhere from 20 to 40 homers next year.  Reggie Willits, who we have less data on, will hit 0 to 3.

By “work”, I propose this test:

1. Take your top 30 young player hitters forecasted (by wOBA or OPS or whatever), all with at most 100 career MLB PA going into 2009.

2. For each player, find his “twin” in terms of BA/OBP/SLG (and position class if possible: C, IF, OF, 1B).  The twin must be under 30 years old, and have at least 2000 career MLB PA going into 2009.

At the end of the season, find out where each of your 30 young players ended up in your percentile forecast. 

And do the same with the 30 veterans.

They “work” if the “clustering” of each group is similar.  They do not work if the young players cluster at the extremes far more than the veterans.

Furthermore, we’d like to see 15 of the 30 players in each group be between the 25th and 75 percentiles.

etc, etc, etc

My guess is that PECOTA would fail this test.  I haven’t looked at Rally’s, but based on how he talks, he’d probably beat PECOTA at this little game.

There is a lot of ripping on BP at this site, and I guess I’m not mathematically advanced enough to understand why.  I’m relatively new to this, and was introduced FJM-->BP-->everything else.  I can get the problems with the replacement level of WARP (as it seems to be a solvable problem), but there seems to be constant picking.  FRAA and WXRL and PECOTA are constantly targeted, and I can’t figure why, when we are trying to measure things that can’t entirely accurately be measured.

Is there no value in the Wieters projection?  What else could be done?  And if those changes were made, what advantages would we have?

Colin is forgetting about park factors, he’s only looking at league factors.  Bowie is a hitter’s park, but he still is correct for the most part.

Andy #6/ I think you are thinking about this with the wrong mindset.  I’m wrapping up a Ph.D. from some prestigous university in the states in Chemical Physics, and many of the things we measure we can’t measure entirely accurately (and trust me, we really try to measure these things accurately), so you approximate certain things and refine as you come closer to getting the right answer, with better methods for explaining what you observe.

The point is to make progress in understanding the world around you, so you can better both understand what happened and what is going to happen.  Now in practice for baseball, as an average fan his PECOTA rating being more accurate won’t change your life at all, but if their going to “publish” their data, we’d like it to be as accurate as possible, particularly since their an important wide read entity, and this could be made more accurate and they haven’t done it is the point being made (I haven’t looked into it, but it sure sounds like Tango is almost certainly correct).  There is some value to that projection, but there’s more value in getting it right.  Clearly for a GM of a team, there is more practical value in getting it right then for a fan. That’s another matter entirely.

That is a really good and thorough article by Colin!  It blows BP’s translations for the Eastern and Carolina Leagues out of the water.  Based on that, you would have to put all of their MLE’s on the suspect list.

I have component Park and League factors for each minor league based on component data from 06-08.  The Eastern league is NOT that much of a pitcher’s league in AA.  The factors for s,d,t,hr,bb,so and errors are:

s .99
d .99
t .99
hr .99
bb .98
so .99
errors 1.01

And yes, Bowie is an overall hitters’ park within the League.  These are regressed park factors:

s .97
d 1.00
t .77
hr 1.06
bb 1.02
so 1.06
errors .97

That being said, I have a projection for Wieters (league and park neutral) that is quite high:

OBP: .402
SA: .541
OPS: .943

As I said, that is park and league neutral.  BAL has an (regressed) “OBA fac” of .99 and a SA factor of .98, so that would make the above:

OBP .400
SA .536
OPS: .936

I don’t know OTTOMH the adjustments for the AL versus a neutral league, but, say it is .98 and .98.  That makes the above:

OBP .392
SA .525
OPS: .917

That is more optimistic than Chone, ZIPS, THT, and Oliver, especially in the SA department, but still less optimistic than Pecota, which is .395/.546.

There is a lot of ripping on BP at this site, and I guess I’m not mathematically advanced enough to understand why. 

The “ripping” is always supported by evidence.  To answer your question: I don’t look to question BPro more than other sites, other than they happen to be in my face alot.  I get on Bill James’ case too, and others. 

Think of me as the Don Rickles of sabermetric comedy.

I’m relatively new to this, and was introduced FJM-->BP-->everything else. 

Well, thanks for coming over then!

I can get the problems with the replacement level of WARP (as it seems to be a solvable problem),

Not only solvable… but solved!  In BP2009, Clay made the long-awaited changes.  And he said that my constant badgering played a part in making him take a look at it.  And he accepted what the rest of us saw.

If we didn’t do what we did, BPro would be constantly giving out those ridiculously high WARP.  Doesn’t that bother you, that BPro authors for years kept quoting WARP, even though they know of the evidence against it? 

but there seems to be constant picking.  FRAA and WXRL and PECOTA are constantly targeted, and I can’t figure why, when we are trying to measure things that can’t entirely accurately be measured.

They are targetted for their faulty assumptions and processes.  FRAA first of all has already been renounced by Clay as well in BP2009 in favor of a fielding system based on play-by-play data. 

Once again, what are we supposed to do?  Say that FRAA is doing the best darn job it can, and not comment that other systems are light years ahead?

I don’t think I talk about WXRL, but I suppose you mean LEV.  Yes, BPro has it wrong.  I provide evidence, and they don’t. 

You should be upset with them for providing substandard measures and not with me for saying that they are providing substandard measures!

And PECOTA is picked on for:
a) the outlandish claim that it is “deadly accurate”
b) for percentiles that have never been publicly tested

Indeed, BPro has skated pretty well scott-free on these issues for a very long time.  They do so by not publicly engaging the analyst community.

And I also say nice things too.  I said elsewhere that RAR and EqA are good metrics.  I love the EqA reports that Clay provides.  His playoff odds report are alot of fun.  Plenty of good stuff.

Is there no value in the Wieters projection?  What else could be done?  And if those changes were made, what advantages would we have?

I don’t see how Wieters forecast is possible, be it from PECOTA or MGL.  First of all, they have to agree that the uncertainty around Wieters must be greater than with Wright and Pujols, correct?

If you start Wieters with a mean close to anyone not named Pujols, then the claim is that Wieters upside (say 75th percentile), as a rookie, will be greater than every non-Pujols hitter in MLB in 2009 begs a scratch of the head doesn’t it?

If the mean forecast for a top hitter is say a wOBA of .400, and we are 95% sure his true talent level is .390 - .410, what would Wieter’s wOBA estimate be?  That his true talent level as a rookie is .370-.430?  I mean, wow.

This is similar to when Nate talked about the chance of Chipper hitting .400.  He used a faulty method, and it produced outlandish results.  We came out right away and provided the evidence for why he was so wrong.  Why no mea culpa on that one?

I don’t see how Wieters forecast is possible, be it from PECOTA or MGL. 

Why?  Of course there is greater uncertainty in that forecast than, for example, A-Rod or Pujols. 

All I am doing is taking his minor league stats from last year (I don’t use single A, but thankfully, his single A stats were great too) and converting them to an MLE.  There is uncertainty in that, but that only contributes to the overall uncertainty of the forecast - doesn’t change the mean forecast all that much.

So now I have a player who has the equivalent of a gargantuan performance in 250 PA in the major leagues.

Now, I regress that - actually I regress the components based on those 250 PA, his age, and his height and weight. The regression is based of course on the proportion of players in the population of professional baseball players who have a true talent of X.  Obviously there are SOME players who have a true talent OPS of .920 or so at age 22 or 23. There are players who have a true talent of much higher than that.  The fact that his sample OPS was off the charts suggests that his true talent OPS is top tier.  That is how the regression works, right? It is just a shortcut for doing a full Bayesian computation (what are chances that a true talent X would hit A in 250 PA and what is the proportion of true talent X players in the population of all players with that body type and age, what are the chances that a true talent Y would hit B in 250 PA, etc?).

Then of course I age adjust at some point (before or after the regression) since we expect his true talent to go up quite a bit from last year to this year, so if we estimated his true talent for last year, it would be higher for this year.

OK, I did all that and granted it is not perfect, but it is correct.  So how and why can you say that these projections are not correct?  How would you know that without doing the computations yourself?

That would be like saying that of 10,000 projections that were done correctly, 10 of them would be for players who were off the charts great players, but you would automatically not believe them.  That makes no sense.  If you are doing your projections correctly, a small percentage of players would be projected to be off the charts, and correctly so.  You can’t automatically not believe them, unless you can show some flaw in the projection system.

Again, yes, the uncertainty is much larger for a player where the projection is based on 250 PA, but if the projection is done correctly, which I assume all of these systems do, more or less, than the mean forecast should be just fine.

I’ll take the average forecast from all the good projection systems, give me a modest 20 points in OPS, and I’ll wager any sum of money on the over for Wieters.  I am asking for 20 points because you are so sure that the projection is wrong…

I don’t mean to poke at you or defend BP, merely to understand.  In all honesty, I think the absolute best sabermetric analysis is done here and at FanGraphs.

However...I also have an irrational love of Joe Sheehan’s writing.  My eyes start to glaze over when I view huge bunches of data, or analysis of PitchFX and the like.  I appreciate that stuff, as without it my sabermetric fanhood would not exist.  But I also really appreciate more standard baseball journalism that uses these metrics rather than RBI, even if it doesn’t add anything new to the debate.  It’s just more my speed...since I don’t have anything to add to the debate either.  I just like to learn what I can, and I guess the opinion-type pieces are more accessible to me.

Anyway.  I seem to be getting off track.  I must first admit that I have a copy of The Book but have not read it yet.

I’m just jumping in here because it wasn’t until recently that I could consider that any of their stats might be wrong, as I can’t really analyze any of them on my own.  I read Clay’s piece in the back of BP09, but didn’t know the source of the distress and couldn’t understand why he was changing things (having only read his side of things) until I came here and figured out the big problem with WARP.  I buy it.  Just keep in mind, I count on you guys (and Seidman, Cameron, etc.) to show me this stuff

I’m not really sure what to make of wOBA.  Do I consider it a less complicated EqA?  What does the sabermetric community at large think of Win Shares?  I always thought Eq stuff was better as it wasn’t ex post facto - it gave real values for what an individual did and didn’t giveth or taketh away depending on strange things that may happen (i.e. the Twins success with RISP last year).  But am I in general in the minority?

I guess I might mean LEV when I say WXRL.  I don’t really know LEV, but I saw criticism of it here and figured it must be part of what WXRL (or SNLVAR) is derived from.  I can see what a criticism might be, as it more measures how a reliever is used than strictly what his abilities are.

I don’t even know any more.  I suddenly find myself buried in statistics.  I don’t know where to start or which way to jump.  I feel like I’ve gone from the far left to the socialist left in terms of the blogs I read, and I’m not entirely comfortable without starting over and reading everything ever written on the subject.

Obviously there are SOME players who have a true talent OPS of .920 or so at age 22 or 23. There are players who have a true talent of much higher than that.

Has anyone ever been forecasted with an OPS of .920 with 0 prior MLB PA, and 250 AA/AAA career PA?  So, not only that, but then you have such a large uncertainty around that mean that it could conceivably be a .960 OPS as the true talent mean?

So, my first request is to present Wieters’ mean wOBA with an uncertainty range for his true talent level.

And you are right that this is a Bayes problem.  The forecasts presented is already giving him such a high certainty to be at the right-end of that talent distribution.

I haven’t heard that since Daisuke Matsuzaka came over.  And since Phil Hughes.  And…

To put it another way: if Wright and Braun were available on the open market (presuming all are average fielding 3B) today, knowing what we know now, would you recommend that Wieters be signed to a comparable contract? 

Basically, I’m very very skeptical that you can have both such a high mean and a high uncertainty.

What would this guy’s MLE have been:
http://www.thebaseballcube.com/players/R/Jim-Rice.shtml

And this guy:
http://www.thebaseballcube.com/players/D/Andre-Dawson.shtml

My guess, total guess, is that if you take all the players who’ve been forecasted with a mean wOBA of at least .400 (MLEs only, no prior-MLB experience), and then you find out what they actually hit in their rookie year, you’d have a shortfall.  A big one.

Well, I don’t think you’ll ever be able to come up with a sample size large enough to test that proposition with any certainly whatsoever.  As I said, if the model is right, then that kind of forecast should be accurate.  If your assertion is true, Tango, we should see more forecasts for rookies like that. We don’t.  That is by far and away the highest rookie forecast I have ever seen. Surely you would agree that eventually there will be a player who has such a high true talent level that he would be forecasted at a .920 OPS in his rookie year.  If the model produces those at about the same rate as they should appear, then you should have no problem with a forecast like that.  Without knowing, A, how often a player with that kind of true talent as a rookie, should come along, and B, how often our basic minor league projection models come up with one, how can you say whether you think this forecast is a good one or not?  You can’t!

And I don’t think that the uncertainty has virtually anything to do with the mean forecast.  Tango, if you are so certain that these forecasts are wrong for Wieters, take me up on the bet for $100 to charity.  I’ll only ask for 15 points in OPS or 10 points in wOBA!

Assuming that Bonds, Pujols, and A-Rod (and in the old days, Bagwell and Frank Thomas) are pretty much the limit of true talent, what would have been their true talent at age 22 or 23?  That should be your guideline.  If any of these projection systems occasionally (like once in 10 or 20 years) forecasts a minor league player at or a little below that number, that should be perfectly reasonable, even mandatory.  If they do that more than occasionally, then yes, there is something wrong with the forecast methodology.  If they occasionally forecast a player much higher than what those guys (Pujols, et al.) would have been or were true talent-wise at age 22 or 23, then there is also probably something wrong with the methodology. Is a .920 OPS at age 22 or 23 out of line with a once or twice in a generation hitter like Pujols or A-Rod? If yes, then we can talk.

But without knowing all of that, it makes no sense to me to assume that a forecast is wrong because it projects a player to be that good, unless we already suspect that the methodology in that forecast system is weak or wrong in the first place.

I mean, by definition (of a good forecast system), the likelihood of someone NOT that great, with a true talent at age 22/23 substantially less than .920, doing what he did in the minors is small!  You would have to agree with that, no? That is why we project him that high!

MLEs by MGL 2001-2003:

http://tangotiger.net/MLE.html

Adam Dunn, who seems to be a good comp here, was forecasted as +59 runs per 500 PA, which is an insanely high forecast, even higher than Wieters.

Jeremy Reed (!!) was +44 runs per 500 PA, which is higher than Wieters.

Half way down we see Miguel Cabrera at +34 runs per 500 PA, which is about where we have Wieters.

So, yes, I think I have a right to be skeptical here.  I think Cabrera, highly touted off the bat, and even younger that Wieters, a guy who has been sensational as a hitter, could be forecast for “only” +34, then no way could we expect Reed and Dunn to have been forecast that high.

I was curious enough that I thought I’d take a guess.  I entered Rice’s and Dawson’s stats into the CHONE system and ran a projection, pretending that their AA and AAA seasons happened in 2007 and 2008.

Some limitations:  I have NO park and league difficulty data for that time whatsoever, so I’m pretending they are equal to equivalent parks today.  For the regression portion of the system, I’m also using late 2000’s data, though if I had been doing this in 1974, I obviously would have used early 70’s data, which with offense being lower, would lighten these projections a bit.

For Dawson I substituted Connecticut of the Eastern league for his AA stats, and Colorado Springs of the PCL for his AAA stats.  For Rice, Pawtucket is still there, I used New Britain for AA.

The results, caeats aside:

Rice 290/349/493 Actual 1975 309/350/491

Not bad at all

Dawson 283/344/482 Actual 1977 282/326/474

Damn.  You picked good young players who did EXACTLY what they were supposed to do.  Give me someone harder.  Like Hanley Ramirez - no way anyone could have seen that coming based on minor league stats alone.

Here, Adam Dunn:
http://thebaseballcube.com/players/D/Adam-Dunn.shtml

It looks remarkably like Wieters.  And, MGL had him forecast as insanely high as per link above.

Retro-jection on Wade Boggs, 1982, after 4 years of minor league data:

314/401/413 Actual 349/406/441.

Not terrible, but some things never change.  Projection systems don’t know how to tell true high BABIP guys from the lucky ones.

Jeremy Reed:
http://thebaseballcube.com/players/R/Jeremy-Reed.shtml

***

Remember, all I did was take MGL’s top 3 hitters from 2001-2003 MLEs, and all of them had high MLEs, with 2 of them insanely high.

That Jeremy Reed looks extremely instructive.  His MLE was like, what .450 or something (can’t do it now, gotta give my kid a bath)?  And the uncertainty level was super high, meaning it was .400 to .500?

Crazy.

The max I can accept is a Miguel Cabrera forecast.

Tango, those are NOT forecasts!  Those are the raw MLE’s!

Are those forecasts?  It doesn’t look like it to me, they look like straight, unregressed MLE’s.  In Reed’s case, MGL is looking at only AA through 2003, which is 242 AB of 409/474/591 hitting.

I’m going to assume that MGL would have regressed that heavily (does he ever miss a chance to regress?) and given him a projection of less than +44.  If Reed had 3 full seasons of .409 hitting in AA, then he might be worth close to a +44 projection.  But he never hit anything like that anywhere else, except for a 58 AB callup in 2004.

All in all, that is a pretty good list that stands the test of time, even if they are not projections.  Most of the LWts are close to what the player would do, at least for those who weren’t old for the leagues.

With Dunn this is based on 350 AB in 2001 between AA and AAA, so the same argument applies.  I’m assuming the table is based only on minor league data, so Dunn’s less impressive performance in the majors from 2002-2003 is not in there.

Here are the top 10 MLE’s from Tom’s list of my 01-03 MLE’s.  To turn them into a forecast, I simply used (in my head) an approximate regression of 50% for 500 PA, etc.  Here is what I get:

Dunn 30
Reed 15
Wilkerson 8
N Johnson 12
Cust 15
Hafner 14
Bay 9
Rios 8
Mohr 8

Average of 11.9 per player

Here are the actual lwts in their first major league season in which they had at least 300 PA (7 were in 04, 2, Reed and Johnson, were in 05, and one, Cust, was in 07):

Dunn 34
Reed -12.5
Wilkerson 16
N Johnson 31
Cust 38
Hafner 42
Bay 30
Rios -10
Mohr 16

Average of 18.5 per player

I assume that I would substantially increase the MLE forecasts for aging, so those 2 numbers (12 and 18) are pretty darn close!

I don’t know exactly what that means, but Tango, bringing up my list to prove your point???

If you want to make a point, use Dunn, since that is the only one with a top tier forecast from my list!

"First major league season” after 2003 I should say.  I did not realize that some of them played in the majors prior to 04…

MGL:  I think what makes your projection surprising is the relatively small amount of regression it seems to imply.  Let’s say we think his AAA performance translates to a 1.000 OPS in MLB.  How much do you regress based on 250 PA—maybe 60%?  (if 500 PA = 50%).  If we regress to league average—which still implies a very good player, given his current age—that would give you a talent estimate of about .850 OPS.  So your projection of .943 suggests either that you are regressing him toward a very high level, and/or projecting a very large talent gain in one season.
Can you walk us through (roughly) how you get to .943?

Thanks for correcting me that the list was raw MLE, which I should have figured.

Note to self: don’t think sabermetrically while preparing a bath.

Note also that I broke up the list of MGL’s players based on OF / IF / C.  So, Miguel Cabrera is at the top of the IF list (half way down that page).

***

I think Rally’s points stands here.  If MGL applied a heavy regression (adding 500 PA of league average performance) how in the world does Wieters get such an insanely high forecast on only 250 PA?  If Wieters, post-regression has a .420 wOBA forecast (or whatever it is), than pre-regression his MLE would be .580 (250 PA of .580 and 500 PA of .340 is .420).

***

The other point is whether you want to regress to league average, or to below that.  After all, rookies who make MLB are not league average hitters, but below that.  So, Wieters is being drawn from a pool of say .320 wOBA hitters, not .335-.340.

(Then again, Wieters was a high draft pick.)

***

So, yes, I’d like to see how MGL can get Wieters from his MLE to his forecast, and end up still so very high.  Perhaps though I should lock in my bet before MGL sees that he made an error.

I asked about Weiters’ PECOTA at a BP book signing, and I spoke with Clay Davenport afterwards about Wieters and projections in general.

Here are points I made concerning Wieters that Clay did not disagree with.
1. Because Wieters played a full season, PECOTA used little or no regression to the mean.
2. Also apparently due to a full pro season, it did not appear that PECOTA used any previous seasons, which for Wieters would be college. They have used college, such as for Pedro Alvarez, but I believe only when they have to.
3. A single outlier season should expect to lose half it’s value for the next season

Oliver’s MLE for Weiters
2005 285/361/450 356 (College)
2006 272/361/461 360 (College)
2007 275/366/450 358 (College)
2008 331/411/561 416 wOBA

I’m not going to only rely on 2008. The college stats show a guy hitting 275-285, 35 DO, 20 HR, 70 UBB. Use diminishing weights on the past seasons, and throw in regression for a 22 year old topping out at AA, and you get
2009 298/376/495 377

The weighting and regression took back 60-65% of the level of play that 2008 was above his college records

I have Wieters raw (but park and league adjusted) MLE’s, with no regression, at .342/.425/.600, which is 1.35 times what I define as major league average (so roughly a 135 OPS+) and around a +56 lwts!

Looking at my projection algorithm for MLE’s, I actually use 250 PA as the 50% regression point (that is probably too liberal - I think I use the same thing for MLB players, which is probably correct).

I regress to major league average for that age (height and weight as well).  I don’t think whether he is a rookie or not matters.  I did some research a long time ago which suggests that age and not experience correlates with performance in MLB.

Now, here is on caveat, which David Gassgo turned me on to last summer:

When you are projecting a minor league player, there are two completely different ways of doing the regressions:  One is to regress to an average minor league player of that age (more or less) and the other is to regress to a major league player of that age. Obviously you will get 2 different results, especially for a lot of regression.

Technically, the former is correct, if you are a team and you want to estimate true talent to decide what a player will do if you bring him up strictly because of the numbers.  However, the second one is “correct” (even though technically it overstates his true talent estimate) if you are already assuming that a player is going to play in the major leagues.

I hope that makes sense. It is a difficult concept to wrap one’s arms around, but it is important.

If we want to use our projections to compare to players who actually get PA in the majors, then we want to use the latter projection (where we regress towards an average MLB player).  If we want to estimate true talent, period, then we want to use the first method, which will yield a lower number.

So you have to define the question before you can put out a projection, especially for minor league players.  What do we expect him to do GIVEN that he is going to play in the majors (we already established that he is part of the population of MLB players) in the majors, or what would he do theoretically in the majors, if we forced every minor league player to play in the majors (which is an estimate of true major league talent)?

Interesting, huh?

Anyway, my projections use the latter assumption.  Obviously if you use the former assumption, you can never test your methodology…

I certainly don’t have a problem using past seasons.  I only use AA and AAA. However, I would also question the importance/weight of college numbers, not the least of the reasons being the accuracy of college MLE’s.

I would love to see your correlations, Brian, between college and MLB numbers.  They would have to be “out of sample” correlations, which means that you can’t use the same sample of data that you used to construct the MLE’s in the first place…

We’re not too far off on the MLE.  I have him at 343/421/548 for Bowie.  But 279/357/416 for Frederick.

Not everyone uses data below AA, but if he really is as great as his AA stats, I wonder why he didn’t hit even better in the lower league.

I’ve got the low projection on Wieters, but even if he hits his CHONE projection he’s an allstar, about as valuable as Russell Martin.

Technical note:

which is 1.35 times what I define as major league average (so roughly a 135 OPS+)

OPS+ = OBP+ plus SLG+ minus 1

In this case, this means OPS+ = 170

***

Marcel uses 240 PA as the 50% regression point for MLB players, so MGL and I are in-synch here.

***

No way can you use the same for minor league players.

***

MGL has Wieters as +56 runs (per 500 PA I presume) as his unregressed MLE.

Lo-and-behold, he had Adam Dunn (a very similar hitter by the way in terms of all those walks and K, plus the amount of PA to begin with… couldn’t have asked for a better comp) at +59, also unregressed.

MGL has his regressed performance at +30. 

So, Wieters falls in the same idea, in that MGL regresses him 50% toward the mean because he has 250 minor league( that MGL used) PA.  So, that’s +28 runs per 500 PA.

+28 / 500 = +.056 runs per PA

Wieters = Dunn.  No problem at all here.

Multiply that by 1.2 to get it into wOBA scale, so that’s +.067 wOBA per PA.  Set league average wOBA at .333, and he comes in at .400 wOBA (which implies .400 OBP, .484 SLG).

MGL is saying this:

OBP .392
SA .525
OPS: .917

Which is not the same thing.  Of course, it depends on MGL’s baseline of what average is.

First thing I’ll point out is that we *should* use the same baseline scale, which is why it always helps to set the league average to zero.

So, I’m going back to Wieters as being +28 per 500 PA.

Adam Dunn, in his rookie year, was +17 in 286 PA, or +30, so MGL nailed it.  In his career, he’s +23.

I would say that:
1. Dunn was probably pretty lucky his rookie year

2. Dunn has had an extraordinarily good hitting career

3. If Dunn was forecast for +30 per 500 in his rookie year, then he would have been forecast for +32, +35, +37, +40 for his following 4 years

4. It’s hard to believe that Dunn, with all those 40 HR and 100 walk seasons could possibly have hit better… he’s have to be the league leader in BABIP as well

5. It’s more reasonable to think that the MEAN forecast for Wieters would be Dunn’s career average (+23 runs), with subsequent forecasts of +26, +28, +30, than to start him off as a rookie at +30, and progress him higher from there.

So, having not done any MLE, and going on my sabermetric guts, I’ll call it: +23 runs per 500 PA for Wieters.

In wOBA terms, that’s .388, if the league average is .333.

***

I’m not that much different from MGL (who is calling him at +28). 

Dunn’s career is .381/.518, with a .383 wOBA (and +23 per 500 PA).

If MGL knocks off 10 OBP and SLG points from Wieter’s forecast, I won’t have a problem.  PECOTA though would have to knock out 15 and 30 points, respectively.

Brian I think has the most reasonable estimate.

A question about MLE methodology and age adjustment for those of you who calculate your own MLE’s and projections.  When you calculate a transition rate from one level to the next to arrive at an MLE aren’t you comparing the statistics that a group of players compile at one level and comparing them with the statistics that they compile at a higher level in the following year?  If so, isn’t an aging component already automatically built into the MLE?  And with a player like Wieters where you are calculating an MLE for a 2 level jump from AA to the majors by chaining from AA to AAA and AAA to MLB, wouldn’t the players in your transition groups be at least 2 years older?  If an aging component is already built into the MLE then isn’t it redundant to also apply an aging factor to the regressed MLEs when calculating a projection?

Peter, it depends on whether the age transition is built in to the MLE in the first place, of course.  Technically, the regression is already built in as well!

Personally, I construct my MLE’s from same year data (players who played in the minors and the majors the same year), so I don’t have the one year aging factor built in.

The other thing, Peter, is that the average age of a minor/major players is maybe 25-26 or so (in the year before they go to the majors), so that you have some players getting better and some actually declining, so if you are using MLE’s from one year to another, it is only capturing a small aging effect, as a group.

Anyway, MLE’s are really a stab in the dark (determining the coefficients, the park and league factors, imbalanced schedules, etc.) to tell you the truth. There is no way that the MLE’s and then projecting a player with 200 or 300 PA is robust enough to distinguish between a +23 and a +28 player.

It’s funny, I actually already have a bet with my roommate: I have the over on Wieters OPSing over 850, conditional on 300 PA for Wieters, which he will almost certainly get if he isn’t terrible in the majors.  This discussion has made me feel pretty good about that bet.

Here’s basically what I do

1. calculate park factors within each league
2. normalize batters by that pf for each league
3. take everyone who played in the Carolina League, and MLB (with 2.5 or more PA/G in a season), created matched pairs, and calculate league to MLB factors for each component
4. normalize batters by league factors
5. take sum of all batters in Carolina League, and apply that league’s MLE factors to find what a typical player in that league would do in MLB, and use that as the regression baseline for each component
6. sum all the 22 year old player’s MLEs, and compare to same player’s unregressed results in step 4 when they were 23, calculating a fudge factor for each component
7. assemble components into a bating line

I don’t chain, I use direct matching of minor leagues to MLB, so yes, much of the age factors for younger players are built into the league factors.

Brian - So in step 3 when you are selecting the matched pairs for Carolina league and MLB, the player doesn’t have to play in MLB in the very next year after he plays in the Carolina league, it could be 2 or even three years later.  Is that correct?

36/Peter - yes. Not saying that I will use this method forever, I am evaluating and looking at other techniques, but so far it has worked pretty well. The most bothersome result I’ve found so far is I do seem to underestimate prepeak players, and overestimate postpeak, so (presumably) the age corrections aren’t working as well as I would like.

Is it really an improvement to break translation factors down by specific minor league?  I don’t see any inherent reason why the PCL should be any different than the IL, given that they are both AAA leagues.  Sure, they have a different composition in terms of environment, but that’s what park factors and league average are for (recognize that park factors are hardly foolproof).  The strength of talent might vary, just based on which ML organizations have affiliates in each league, but that’s a moving target. And I also recognize that the difference between A and high-A is not exactly formal.  Still, I question the utility of cutting your sample size in half by drawing distinctions between two leagues at the same level.

I’d be interested in hearing why I’m wrong from those of you who are in the projection game.

38/Patriot - the easiest way to do park factors is to make each set of parks (a league) mean to 1.00. Thing is, a 1.00 HR park in the PCL isn’t the same as a 1.00 HR park in the IL, in this case mostly because of altitude. You could very easily have one league with a bunch of small parks, and another with a bunch of large ones.

There’s no easy way to compare PCL parks to IL parks, but I can compare PCL hitters to MLB, and IL hitters to MLB, so I first normalize batters to their league, then compare that league to MLB.

Patriot, he is just doing a roundabout way of doing league factors.  I compute league factors based on components of one league (like INT or PCL) as compared to all AA or AAA.  Then I use the same MLE’s for everyone in AAA or AA, as you suggest.  I think it ends up being the same thing.

Blackadder, if you have a min number of PA for the bet to vest, that is real good for you, as you get the benefit of selective sampling. He has a lousy first 100 or 200 PA and he gets sent down.  He gets lucky and he stays up.  His average OPS for 300+ PA in the majors is probably 10 or 20 points higher than his mean projected OPS!

What MGL said.

Alternative is to say, all of AAA has HR rate of .040, PCL is .045, IL is .035, so bump up PCL pf’s by that amount, and decrease IL pf’s. I don’t think there’s any difference in the answer, I used to use this method, but went with the one which was easier to code. (Normalize player to league, then normalize league to MLB).

MGL: how do you get Reed from +44 unregressed to +15 regressed?

From 2001-03, which is what those MLEs are from, he had 268 PA, which means that you would regress a bit less than 50%.  This puts him at +23 runs per 500 PA.

I think you must have added in his 2004 MLEs?

***

As for the right regression point for all these hitters (i.e., hitters who have never played in MLB, but the teams have decided are good enough to play MLB), it would be whatever the “true” first-year players do, which I presume is something like -0.01 to -0.02 runs per PA.  (We should take the median, and not the average, otherwise Braun and others would have a disproportionate impact.)

Also, there is no way that minor league stats could possibly be as informational as MLB stats, and so, adding 250 PA of regression is too low.

If for example we are adding in 350 PA of -.02 runs per PA (or -7 runs on 350 PA) to Wieters’ +28 in 250 PA, that gives us +21 in 600 PA, or +17.5.

Wieters has something else going for him: he’s drawn not only from non-MLB players about to make it, but a high first-round pick (like Braun, Tulo, Longoria, Jeff Clement).  So, the regression point should be higher than -7 on 350 PA.  Maybe it should be +2 on 350 PA.

So, now we have +30 on 700 PA, which is +21 per 500 PA.

***

After Jeremy Reed’s insane 2003, the Whitesox did NOT bring him up.  Therefore, he would not be part of the -.02 runs per PA group of rookies, but perhaps -.03 or -.04.  However, he was a 2nd round pick, so that gives us more comfort, and that might bring him up to -.02.

So, we add -7 on 350 PA to his +24 on 268 PA for a total of +17 on 618 PA, or a rate of +14 per 500 PA.

***

We see here how Wieters goes from +56 to +21, and Reed goes from +44 to +14.

***

We could have used even more information, like Reed is a very small guy, and Wieters is not.  There are lots of different regression points we can use.

***

The key though, and I’ll blast PECOTA here, is that you MUST, absolutely MUST know the number of PA in your sample.  My guess is that Nate makes zero distinction between whether Wieters had 250 PA or 500 PA in AA.  And that’s why it’s not a good forecast.

"My guess is that Nate makes zero distinction between whether Wieters had 250 PA or 500 PA in AA.  And that’s why it’s not a good forecast.”

Well, I think Colin’s original point about the MLEs plays a role too.  Compare Rally’s and BP’s MLEs (OPS):
BP/Rally
A 1.024 / .773
AA 1.085 / .967
Average them, and BP is saying Wieters was already a 1.054 player last year, while Rally is saying .870.  Even if you can agree on an amount of regression, that will lead to very different projections.

I just asked Nate:

“Whether Wieters had 250 or 500 PA in AA, his forecast would not change”

That is, you only look at rate stats, and you do not look at the number of PA (or IP) of the player you are forecasting.

He responded almost immediately:

That’s right.

I’m really hoping that Nate misunderstood the question. (I’m also starting to wonder if PECOTA is as good as it is on accident - a lot of things about PECOTA are questions for me now.)

There’s nothing to misunderstand.  The question couldn’t have been any clearer.

And, it’s not like I had any doubt of his answer, as it was completely consistent with everything I know about PECOTA.  And it was also consistent with why he was so wrong about Chipper Jones last year.  And it’s also consistent with why the PECOTA percentiles are independent of the number of PA or IP of the player in question.

PECOTA starts with the rate stats, and does not look at the number of PA or IP.  It then finds comparables based on the comps’ rate stats.  It then looks at the total number of PA, IP, G, years played of those comps (and the comps of the comps).

It “works” because most players forecasted are not 22-year old players from AA with 250 PA.  The things that PECOTA does very well is diminished by the things it does that are obviously wrong.  That it does as well as it does is the same reason that Marcel does as well as it does: there’s not much to this forecasting game to begin with.

That probably came off more snide than justified. The more I read about PECOTA, though, the more I start to think that their baseline forecasts are overengineered - the way they handle a lot of things strikes me as bringing out a sledgehammer when a ballpeen hammer would do. (And there’s a lot more ANOVA/regression involved than I would like - I consider regression a tool of last resort.)

"It then looks at the total number of PA, IP, G, years played of those comps (and the comps of the comps). “

I should append to the end: “ in subsequent years”.

The sim scores are only used in generating the “career path adjustment” - it’s the part of PECOTA that gets the most attention, but most of the real work is done in generating the “baseline projection.” Give me a second to look over what Nate wrote about PECOTA in BBTN and the BP 2003 book (hint: go to Amazon.com and use the Look Inside feature; you can read the whole PECOTA article from ‘03 if you’re patient).

Nate doesn’t spend much time talking about the baseline forecast in either book; he devotes much more space to the “career path adjustment,” essentially sim-score generated aging curves. I think this is a shame - the sim-scores thing is interesting, but BP sells it as the core of PECOTA, when really the baseline forecast (IMHO) is the important part.

Nate starts off where all of us start off - a weighted average of the past three seasons. He regresses to the mean, he says.

This is where it gets tricky - he doesn’t provide very good examples, but if I had to guess at what he’s doing here, this would be my guess:

He breaks a player’s batting line down into components (this is probably the unclearest part of the whole thing - what components is kind of a grey area). He then uses some variant of multiple regression to predict the next season’s batting line, based upon several components. For example, if he’s predicting hit rate, he’d look at several variables - prior hit rate, strikeout rate, some kind of speed score, etc. For predicting home runs, he’d look at not only home run rate, but doubles rate, etc.

This is the baseline forecast that is then fed into the career path adjustor. And playing time IS a factor in the career path; it’s one of the factors used in determining sim scores.

If PECOTA’s making some mistakes, doesn’t that mean it has the potential to be even better by tightening things up?

Past number of PA is NOT a factor in determining the baseline forecast.  Wieters could have had 250 PA or 2500 PA, and as long as he had the same rate stats, it’s off those rate stats that he’s being judged.

And similarly, his comps are matched on their rate stats.

The past playing time may be a factor in determining the quantity of future time (as it should be).  But, it MUST also be a determination in the current baseline forecast.  And it’s not.

Sky, as I said:

The things that PECOTA does very well is diminished by the things it does that are obviously wrong.  That it does as well as it does is the same reason that Marcel does as well as it does: there’s not much to this forecasting game to begin with.

So, yes, it could be better.  But no one’s been listening to me for five years on this issue. I’m like Andy Dufresne trying to get library books.

So long as you’ve got Nate’s ear, Tom, have you asked him about the DTs used in the Wieters projection?

I find this hard to believe.  So if a player gets 5 hits in 10 at bats, he would project as the greatest hitter ever?  Is that what we’re saying here?

Or is PECOTA avoiding an extreme problem like that by refusing to project players without a certain minimum PA?

Guy.

In #43 you are listing the actual OPS Wieters had, not BP’s mle.  From their site I found these mle translations:

AA 1.029
A+ 0.833

So they aren’t discounting the minors as much as I am, but it’s not that far off.

I would presume he’s got a minimum baseline, like 200 PA or something.

As for Nate’s ear, how about you guys come up with a series of questions, and then I’ll submit them to him.

Rally:  thanks for correction.  But looking again at Wieter’s Pecota card, I see the translations as:
AA:  1.063
A+: .909
I don’t know which of these translations, if either, Pecota actually used.  (You gotta think even the BP guys have trouble keeping all this straight at this point!) But these #s are still pretty far from yours.

1. Are there a standard number of PAs of regression given to each batter?
2. Looking at the season stats on a PECOTA card, I can see that each is regressed (those with smaller PAs are closer to the mean) - for the projection, are they regressed and then summed, or summed and then regressed once?
3. Are college stats routinely used, or only when there’s a lack of pro stats on a popular player?
4. Are the league factors recalculated yearly, or do they use long term trends?
5. Could old PECOTAs (before 2006) be made available for accuracy testing?

Guy - According to an e-mail I got from Kevin Goldstein, the ones on the PECOTA card are the ones used in the translation. BP updates their park/league factors at the end of the season, and it’s that update where I think things went awry.

League factors should be (and this could have changed since Clay wrote about it last) based upon five years of stats.

I am assuming that number of PA are used in the similarity scores, so they are implicitly used in the forecast.  If a player has 10 PA and gets a hit in 5 of them, I assume that the 5/10 performance will have little bearing on his comps.  In fact, his comps will be almost all players even though he only has 10 PA and was 5/10 (which is a rare feat, having only 10 PA in a season and going 5/10 to boot).

Bad assumption MGL. 

While you may “hope” that a guy with only 250 career minor league PA would be compared to other players in the same boat (Dunn, Reed in 2003, etc), this is NOT what Nate is doing.

Re-read post 44.  The number of PA is irrelevant to the comps.

Things that this thread and Colin’s article have me thinking about

PECOTA could have a bad league factor for the Carolina league. I will compare other player’s 2009 PECOTAs who were in the CL to Oliver and see if there’s a general shift.

They could also be relying to heavily on comps for the forecast, and then with Wieters get a guy who the system shows with no good comps. In mapping, this is working past your control, a big no-no. Data past the control can flop around too high or low and the system won’t be able to pick it up.

Tango, I know you THINK you asked him a clear, unambiguous question and that he gave you a clear, unambiguous answer, but if we use the reductio absurdem argument, which is, “Is he going to forecast a player with 1500 PA with a WOBA of .380 the same as a player with 100 PA of .380, of the same age, body type, position, etc. the same? and will he use the same comps,” the answer is clearly, “No,” so your question and his answer are NOT unambiguous.

Just to piggyback on what MGL is saying here - in other places, Nate has made it very clear that playing time IS one of the criteria used in his sim scores. (Check the Kevin Mass chapter of BBTN for a full list of the components of the sim scores ranked by priority - playing time is number 4.) So this is rather perplexing.

Brian, I can only go off the list of top comps on his PECOTA card - which is not the full list - but a weighted average of them (with the sim score as the weight) showed a 4% decline between y and y+1 in wOBA. The sim scores are NOT what’s causing the forecast to be so far off.

I also want to add, as someone who does projections and a zillion other things in my “spare time,” due to the complexities of these methodologies and algorithms, and the fact that BP does a zillion other things as well, there are bound to be mistakes and bugs all over the place.  Sort of like Windows, or just about any other complex computer program.  Doesn’t mean you want to throw the baby out with the bath water.  I agree with Tango, though, that given the possible and even likely bugs and mistakes in Pecota, it is probably not worth the effort as a halfway decent Marcel type of forecasting system is going to do almost as well.  Maybe without the bugs and mistakes Pecota would do a lot better, but unless you make Pecota your life’s work, you are not going to be able to avoid those…

MGL, if you change your 100 to “whatever his minimum threshhold is”, the answer is yes, even though we all agree it should be no.

***

What is not clear in this question:

“Whether Wieters had 250 or 500 PA in AA, his forecast would not change”

That is, you only look at rate stats, and you do not look at the number of PA (or IP) of the player you are forecasting.

How can you interpret the question in any other way?  Can you rephrase my question in a way to which Nate thinks I may have been asking him?

His answer ("That’s right") is crystal clear of course.

“Whether Wieters had 250 or 500 PA in AA, his forecast would not change”

That is, you only look at rate stats, and you do not look at the number of PA (or IP) of the player you are forecasting.

Tango, regardless of whether a question and answer are crystal clear and both the questioner and answerer are reasonably sane, intelligent people, sometimes there is a misunderstanding or some other “disconnect.”

If I told you that I asked Dewan if Jeter was the best fielder in baseball and he said, “Yes,” even though the question and answer are crystal clear, that does not mean that what was said is true.  It could be that one misunderstood the other, it could be that someone was joking, or it could be any number of reasons that are not immediately evident.

So please don’t assume that just because your question and the answer were crystal clear on their face, that that means that there is only one correct interpretation!  That is absurd.

OBVIOUSLY Nate does not mean or believe that if a player has a wOBA of .400 in 16 PA that his forecast is going to be the same as if he had 1500 PA and a wOBA of .400.  That WOULD be true if the crystal clear second part of your question and his crystal clear answer were true.  It is not. 

Sorry to have to yell, but you are being ridiculous!  Yes, we all agree that the question and answer that you typed into the computer is crystal clear.  But that does not mean that it is true on its face.  It is not.  Nate nor Pecota does not have the same forecast for a player with 3 PA as he does for a player with 3000 PA, even if they have the same rate stats! For that to be true, which I think we are all in agreement that it is, there has to be some disconnect between the second part of your question and his answer!

Just one more thing.  You should not criticize someone as smart as Nate for something he said which appears to warrant further explanation.  If he chooses not to explain himself, which he is under no obligation to do, then you just chalk it up to some kind of misunderstanding…

I never said 3 PA or 30 PA. 

I said 250 and 500 PA.  I recognize that Nate uses a minimum PA threshhold.  I asked specifically about 250 and 500, and he said it didn’t matter.

If I asked about 25 and 2500, I am sure he would say “duh, of course I need something more than 25”.

So, SPECIFICALLY about my question, with 250 and 500 and Wieters, exactly what in that EXACT question could be misinterpreted?

Nothing!  You keep replying to everything except for that specific question!

I have three separate instances in which Nate shows that he is not using the number of PA to establish a level of uncertainty:

1. My direct question about Wieters, which was unambiguous.  You may say that all questions are ambiguous, and therefore, give a Clinton “definition of is is”, but not here.  The question is clear, and if someone wants to try to reinterpret that exact question some other way, please do so.  No one has done that yet.  All I hear are generalities.

2. His analysis of Chipper hitting .400 last year.  Again, very clear what he did there: he did not look at the number of PA he had (219) and simply treated it as if it was a typical season’s worth.

3. His percentile forecasts bear no relationship whatsoever to career PA (or last-3-years PA).  Whether you are a rookie, or a veteran, the range is the same.  That makes no sense.

Indeed, someone as smart as Nate should be criticized more for not explaining himself clearly and not less.  And, yes, he IS under an obligation to explain himself, if he’s trying to sell a product or service to an unsuspecting consumer.

And to establish that position, Nate himself criticizes pollsters just as I criticize saberists. 

I’m like the Consumer Reports of sabermetrics. 

Tango, I am not sure I understand your criticism of Nate’s Chipper article.  What he does in that article is start with some prior about Chipper’s true batting average (which is normal with mean around .310).  Then he simply applies Bayes’ Theorem to update that estimate of Chipper’s true talent.  In particular, he certainly takes into account the fact that it was only 219 PA; if it were a full season, the Bayesian update would be bigger.  Now, you can of course argue that there was something wrong with Nate’s article, but I don’t see think he was so stupid to not take into account the number of PA Chipper had up to that point.

Blackadder: that is not what he presumes.  Here is the thread: http://www.insidethebook.com/ee/index.php/site/comments/chipper_does_not_compute/

Here’s what I said:

I don’t see how .348 is possible, unless Nate is treating his 219 at bats as if it was 500-600 at bats.

Please put your comments in that thread after reading everything in there.

Tango, we are kind of talking past one another and I don’t fully comprehend your “argument” with Nate, but you are missing my general point about questions and answers.

I am not arguing that your question and the answer were not crystal clear on their face. So there is no need to point out that they are/were. Please refer to my “Dewan and Jeter” example.

My general argument that is even if a question and answer are crystal clear and unambiguous, there is always the chance that the answer is not meant to be what it seems on its face, as in my Jeter example, for any number of reasons, not the least of which is that just because you think something is clear and unambiguous does make it so for everyone. 

And for the record, OBVIOUSLY your question (and his answer) was NOT crystal clear and unambiguous since I interpreted the second part to apply to ALL players and ALL forecasts and I interpreted it to mean ANY number of PA, such as 3 or 3000.  If anything, the way you worded the second part of your question supports my interpretation.  So your two-part question is ANYTHING but crystal clear!

I mean, geez, here is the second part of your question in black and white:

That is, you only look at rate stats, and you do not look at the number of PA (or IP) of the player you are forecasting.

You say “of the player”, so that should mean “any player” and have nothing to do with Wieters, and you say, “you do not look at the number of PA of the player” and you say nothing about minimum number of PA, etc.  You may think that my interpretation of your question is somewhat absurd, and I think that his answer is somewhat absurd.  Therefore there is anything but “crystal clearness” in that question and answer. Anything but…

No, we are not talking past each other.

Suppose that Nate interprets the second part of my question as you are doing, and therefore, has no qualifier, no minimum threshhold, and treats the second part as distinct from the first part (which was all about Wieters).  He answers “That’s right”.  If he gives that clear answer to something that could not possibly have that answer (because if he has just 3 PA, it’s impossible for him to say that), then Nate is a fool.  Nate is not a fool.

So, the second-part of the question can be interpreted as one or both of:
a. there’s some minimum threshhold of PA or IP (which is no higher than 250 PA, based on the first-part of the question that was about Wieters)

b. it was directly tied-in to the first part of the question which was only about Wieters

There is no other interpretation that you can have, that would allow Nate’s answer to be a sane “That’s right”.

I concede that in other situations, not my specific question here, there may be other interpretations.  But, that does not apply in this specific instance.

***

Regardless, pose as clear a question as you’d like, and I’ll forward as-is to Nate, or if you want his email address, let me know, and you can ask him exactly what you want to ask him.

Ask him also why the percentile ranges are independent of the number of career or last-3-years PA.

***

Can someone perhaps post the 90th and 10th PECOTA percentile EqA for Wieters, Dunn, Berkman, Manny, Pujols, and Wright?

Wieters 90: .357
Wieters 50: .318
Wieters 10: .286

Dunn 90: .339
Dunn 50: .297
Dunn 10: .271

Berkman 90: .349
Berkman 50: .322
Berkman 10: .291

Manny 90: .345
Manny 50: .314
Manny 10: .283

Pujols 90: .376
Pujols 50: .354
Pujols 10: .337

Wright 90: .352
Wright 50: .327
Wright 10: .290

Are the Pecota percentiles true talent intervals, expected performance intervals in the upcoming season, or both (observed plus real).

As you know (I don’t have to tell you), the fewer PA a player has, historically, the wider the interval of our true talent estimate.  That is also true for players of different types, ages, injury histories, although the primary variable in the size of that interval is past number of PA.

As two extreme examples, a player with 3 PA will have a league average forecast (for his type of player) with an interval (percentiles) equal to the spread of talent in that same population.

If a player has 10,000 PA, our forecast of his true talent has a very narrow confidence or error band.

The second part is the random variance around the expected performance. The width of that is simply the number of expected PA, although we are uncertain of that as well, which adds some more variance or width.

So, again, I know this has been asked and answered already, but are the Pecota percentiles only the width of the error band surrounding a player’s true talent estimate or do they also include the random variance of performance in X number of PA, given a fixed true talent level?

Either way, the width should primarily be determined by the size of the player’s history (although, as I said, other things contribute to the uncertainty of the forecast, which is all one part of this error bar is).  If the error bars, or percentiles, or whatever you want to call them, also includes the random variance around a true talent mean in X number of PA or IP, then the width also is a function of how much playing time you forecast (as well as the uncertainty of that playing time forecast).

As another example, even though that player with 10000 PA has a narrow uncertainty bar around our estimate of his true talent, and hence his forecast, if we were to forecast him for only 20 PA, regardless of how narrow the error bar on his “true talent” forecast is, the error bar on an estimate of his actual performance is obviously going to be large.

So, again, it really depends on what Pecota means by these percentiles - uncertainty around their true talent estimate, around the estimate of the player’s actual performance in X number of PA, or both.  It is likely to be either the first one or both and NOT just the last one.

Now, if Nate is strictly or even primarily using comps for the forecast, then all he has to do in the percentiles is to see what similar players actually did in the next year and that determines those percentiles automatically.  The problem with that is this, for example:

Let’s say that we have a player like Wieters, who with only 250 PA or so, we have little confidence in our estimate of his true talent, even if we are pretty sure it is really good (the error bar width is still large).  And let’s say that Nate’s comps are players who happened to have performed, in the next season (after the similar season or seasons) in a narrow range.  If Nate is simply using the distribution of the performance of those similar players in subsequent seaons, and nothing else - not taking into consideration the fact that we only have 250 PA for Wieters, other than impicitly in the comps, then he is going to be making a mistake by having a narrow error interval for Wieters.  That can’t be, regardless of the comps.

What if, for example, we has that player with 3 PA. We KNOW for a fact that our error bars around our forecast are about equal to the spread of true talent i the population we think that player came from. But, what if our comps happened to have a narrow range of performance in the subsequent year?  We might be tempted to use that range for our error bars on this player, and we would be wrong!  Of course, for a player like that, the pool of comps should be hundred of players.

For a player like Wieters, the pool of comps is going to be small, and thus, it is likely that their distribution of performance in the subsequent year is going to be narrow.  If we use that as our percentiles for Wieters, we would be making a mistake.

I think that is what Tango (and I) has been talking about for years.

But the issue cannot be discussed and/or resolved with one, two-part question, I am afraid…

I put together this spreadsheet while looking at the 10th, 50th, and 90th EqA percentiles.  First tab “Data” looks at how far apart the 10th and 90th percentiles are for each player (spread) and which side of the 50th percentile has a greater range (skew).  I put notes to explain the headers if you mouse over them.  The second tab “Graph” puts it in graphical form via a high-low-close type graph.

Pujols has smallest spread at 0.039, Wieters the largest at 0.071, though Dunn is at 0.068.  Dunn has by far the largest spread above the 50th percentile, while Wright has the largest spread below the 50th percentile.  Manny’s spread is even.

Hugh: thank you!

As you can tell, the 90/10 percentiles are indistinguishable between Wieters and everyone else, in terms of range.

What they are *supposed* to represent is the chance that we will observe those percentiles this upcoming season.  As a result, it is a combination of:
a) the true mean plus
b) the uncertainty of that mean plus
c) the sampling you’d expect based on 600 (or whatever) PA

a) and c) are the same for all the regulars.  That means b) is much less for the veterans and much higher for the rookies.

That is NOT what Nate does for b).  His uncertainty of the mean is *also* the same for everyone!

***

As for what the Percentiles represent:
http://www.baseballprospectus.com/glossary/index.php?mode=viewstat&stat=2

A player’s Percentile Forecast is a representation of the player’s expected performance in the upcoming season at various levels of probability. For example, if a pitcher’s 75th percentile EqERA forecast is 3.52, this indicates that he has a 75% chance to post an EqERA of 3.52 or higher, and a 25% chance to post an ERA less than 3.52. Higher percentiles indicate more favorable outcomes.

***

If you want to resolve it to your satisfaction (MGL), then write out your questions for Nate.

***

Wieters’ page is available for free:
http://www.baseballprospectus.com/pecota/WIETERS19860521A.php

A 10% chance that he will perform at an EqA of under .286?  That is consistent if he has a true mean of .318, with uncertainty of that mean of ZERO.

SirK: except for Pujols, they are all around the same general range (.060 to .070 spread).

The Pujols one is another sign of concern:
Pujols 90: .376
Pujols 50: .354
Pujols 10: .337

How is it possible that he has only a 10% chance to perform below .337?  One SD for the binomial is around .020 for 600 PA.

Just by luck, he will be .334 or less 17% of the time.  And yet Nate shows him at below .337 10% of the time.  That makes no sense.  Even if he’s 100% certain his true talent is .354, he has a 17% chance he will perform 20 points below that.

Their glossary speaks specifically on the percentiles as the chance we will observe that performance, not the chance that that’s his true talent level.

Aside from the issue confidence levels for low PA players, isn’t there a problem with Nate’s assumption of symmetrical error bands around his talent estimates?  If you look again at his Chipper article (http://baseballprospectus.com/article.php?articleid=7652), his estimates of Chipper’s true talent look symmetrical around the mean forecast.  So he begins with the assumption that pre-2008, Chipper with a .310 average was as likely to be an unlucky .330 hitter as a lucky .290 hitter.  But that can’t possibly be right—while it’s true that a .290 hitter and a .330 hitter are equally likely to hit .310, there are probably 50x more .290 hitters than .330 hitters in the population.  So his distribution doesn’t seem to incorporate anything we know about the larger population. 

This problem gets magnified after he does the Bayesian calculation based on Chipper hitting .420 to start the season.  First, he estimates a new true talent of .348, which you can only arrive at if your prior estimate of his talent ignored the actual distribution of BA talent in MLB.  But worse, he now assumes that it’s as likely Chipper has become a true .368 hitter as a true .328 hitter.  You can only say this if you set aside everything you know about the distribution of BA talent in baseball—the highest active career BAs today are around .330.  And just thinking about Chipper alone, the chance that a true .368 hitter would have hit just .310 over 15 previous seasons—or even 3 seasons—is essentially zero.

And the exaggerated possibility that Chipper was truly a .360 or .370 hitter must have played a big role in Nate’s estimate that Chipper had a 12% chance of hitting .400 for the season.  I’m sure that most of the simulations that had him hitting .400 were those in which he was assigned a very high true talent estimate.

Wieters’ percentiles do look a lot like a ‘skewed’ cumulative distribution function with n equal to something around the 250-275 range (I used the weighted EqA of .320). 

For Pujols, if we take n=600, p=.354, and x=600*.337=202, the probability that x<=202 is around 20% if EqA is treated like a binomial distribution.

I don’t think you can treat EqA like a binomial like that, because of its exponential scale. I could be wrong.

Guy, I don’t know that PECOTA’s percentiles are assumed to be symetrical. They’re residuals of the sim score process, IIRC.

On page 266 of Baseball Between The Numbers, Nate explicitly says that playing time is a factor in the sim scores. That’s why I have a problem with buying into the discussion above.

And Tom, if you would like to pass along a question to Nate for me, please ask him if he’s got an opinion on the Carolina League difficulty rating.

Colin: you can convert EqA to wOBA easily enough, and then use the binomial-like equation shown in The Book.

As for the sim scores, all I can say is that he probably uses that to forecast playing time.  That is, it’s two parts: the rate part, and the playing time part.  He treats the rate part independent of the number of trials.

I don’t know that you CAN convert EqA to wOBA easily enough. EqA is based upon R/O, rather than R/PA. And it’s just WEIRD. Look at the relationship between EqA and R/O:

http://farm4.static.flickr.com/3410/3408145125_4a9c293db1_o.png

I guess you could treat EqR/Out (or EqA ^2.5 times five) as a binomial - although I’m not sure of that either, this isn’t my baliwick.

And that interpretation isn’t supported by anything written about PECOTA’s sim scores by Nate, either in the BP 2003 or BBTN. Nate takes 14 dimensions of a player’s performance - ISO, playing time, speed score, career length, height, etc. - and measures the distance between them, weighted according to a regression (okay, an ANOVA) against performance in y+1.

Then, he compiles that all down into a single unit by using a variation of the Pythagorean theorum (the one from Pythagoreas, not Bill James) in n-dimensions:

http://betterexplained.com/articles/measure-any-distance-with-the-pythagorean-theorem/

(Again - anyone who wants to read more about the sim scores, go to Amazon.com, search for Baseball Prospectus 2003, click on the cover and serach for PECOTA. It’ll all be there, although there’s a limit to how many pages you can read at a time.)

Now, it’s possible that his “playing time” variable isn’t PA, but something like games started or a dummy variable for starter/bench player or what have you. But based upon previous publication about PECOTA, a player’s playing time DOES affect his comps list.

This doesn’t mean that he’s using playing time to figure his baseline projection - it’s pretty certain that he isn’t, and it’s pretty certain that he should be. (And I don’t have a proof, but I’m very confident that the baseline projection is more important than the career path adjustment - using PECOTA’s baselines with Marcels’ aging adjustments would probably be closer to PECOTA than using Marcel’s baselines with PECOTA’s career paths.)

A great read on how the career paths do - and don’t - affect a players’ projection is Silver’s most recent article:

http://baseballprospectus.com/article.php?articleid=8663

People who subscribe to ESPN Insider but not BP can also read the article here:

http://insider.espn.go.com/mlb/insider/news/story?id=4017235

Late to the party, but I have to get a few things off my chest.

“Just one more thing.  You should not criticize someone as smart as Nate for something he said which appears to warrant further explanation.  If he chooses not to explain himself, which he is under no obligation to do, then you just chalk it up to some kind of misunderstanding…”

Strongly disagree.  Especially about the qualifier “someone as smart as Nate”.  We’ve got a lot of smart people on this board, and I have no idea how they’d stack up against each other on a wonderlic test or anything.  Such a courtesy should be extended to all, or none.  In my view no one is immune from criticism.  He’s under no obligation to further explain himself, of course, but if he chooses not too then it’s open season.

I will never assume that just because someone appears to be smart, they must know what they are doing.

I’m not sure if PECOTA percentiles are supposed to show an error around an estimate of true talent, or assume that true talent is known and the percentiles are variation around that fixed estimate for the upcoming season.

To be honest, I’m not quite sure what mine are doing.  I would lean towards estimates of true talent, since the sample size is used to generate the ranges.

I don’t think this is quite correct though:

“If a player has 10,000 PA, our forecast of his true talent has a very narrow confidence or error band.”

It might be if a player could bat 10,000 times in a season, but by the time a player has that many PA he’s at the tail end of his career and much of those PA have little relevance to his current projection.  For example, does anyone really think we have a better estimate of Ken Griffey Jr.’s CURRENT ability level than we do for David Wright?  I don’t.

Guy, don’t forget that in most of the projection algorithms where they do a “straight” regression towards the mean, that regression algorithm also assumes a symmetrical distribution.  Not correct of course, but it may be closer than you think.

The reason it ends up being closer than you think is the selective sampling/playing time issue.  When a player does badly (gets unlucky), he gets fewer and fewer PA, and vice versa if he does well (gets lucky).  So it makes it “look” as if the true talent probabilities around a mean true talent forecast are much closer to being symmetrical than they really are.

So while the forecasts are not good in terms of the accuracy of true talent, they’ll end up looking better once playing time is accounted for.  I talked about that in another thread (actually, it might have been this thread).  Depends on what you are forecasting.  Are you forecasting a player’s true talent or are you forecasting his performance, assuming that he gets to perform or assuming a certain number of PA or IP?  It is usually the latter, at least de facto it is (since we have no way of evaluating a forecast if a player does not play and we don’t really care if he does not play).  Pecota is clearly the latter.

So if you are going to forecast a player with the expectation that he is going to play, and even play more than a few games, and you know that you will only be evaluated if that player plays and sometimes only if he gets a min number of PA, then you definitely do NOT want to use the distribution of talent among all players in that player’s population to do your forecasting.  That distribution will be heavily skewed towards the left.  You want to use the distribution of players who get significant playing time in the future.  That distribution will be more symmetrical since it does not represent “true talent” only.

Of course, if you are working for a team or something like that, you want to do just the opposite!  You don’t want to project a player assuming that he is going to play next year or that he is going to get a min number of PA or IP.  You want to project his true talent in order to make those playing time decisions!  THat requires a completely different, and much more pessimistic, algorithm.

As far as Pecota and the percentiles, clearly there are a lot of problems there.  Again, I chalk it up to everything I said in #66 above. I don’t think that Nate is doing anything “deliberately” wrong.  I think there are just lots of bugs and fundamental weaknesses in Pecota, despite it being a good forecast system. And that Nate does not have the time nor the inclination to fix those bugs or clean up the system…

Colin:  I’m not making any assumption about symmetry in Pecota projections.  In the Chipper piece, Nate specifically tells us is true talent estimate for Chipper, both pre-2008 and then adding in his early-2008 performance.  In both cases, the curves appear symmetrical (or very close).  Now, as MGL says, when we have a lot of data on a player, this may not matter much for our projections.  If Nate gives equal probability to Chipper being a .300 or .320 hitter (if mean = .310), while that’s not strictly correct it may not affect the projection enough to matter (binomial variance is larger at that point). 

But in both the Wieters case, and our Chipper-hits-.420-early scenario, I think this does make a difference.  Even if you think our best estimate of Wieters’ talent is a .920 OPS, it most certainly isn’t the case that .820 and 1.020 are equally likely.  Pecota appears to be saying they are (in fact, it seems to say his upside is larger!).  And the idea that there was even a probability greater than .00001 that Chipper was—and always had been!—a .370 hitter is just silly.

As I understand it, PECOTA’s percentile forecasts are entirely a function of the career path adjustment. In short, a player’s 10th percentile forecast is what the bottom 10% of his comps did in the following season; a player’s 90th percentile forecast is what the top 10% of his comps did.

You can look at Wieters’ comps by wOBA in y and y+1 here:

http://www.editgrid.com/user/cwyers/Wieters_PECOTA_Comps

Coli #90, yeah, that is what I always assumed.  And of course because of sample size issues, you will get lots of implausible results.

I have always said that the “comp” method for forecasting and for showing any kind of likely distributions is perfect and brilliant, that only applies if there is a ridiculously large number of comps and playing time for those comps.  That will never happen.  Therefore you HAVE to use patterns from ALL players in addition.  That is especially true for unusual players.  There is no chance that a 100% or predominantly “comp” system is going to work for a player like Wieters.

The other problem with a “comp” system is you completely get results which answer the question, “If this player ends up playing like similar players ended up playing.” So you completely eliminate the chance to do the other kind of projection, which is, “Not knowing whether this player will ever see another MLB game again...”

If you are using a “comp” system, and you want to get granular in terms of what you are presenting, like “percentiles,” you really need to incorporate some common sense and some “theoreticals” (inferred from ALL players) into that presentation, otherwise you are going to get all kinds of flukey results because of sample size issues.

Imagine this:  Let’s say that you also wanted to project a player’s aging curve.  Pecota does this somewhat in their long-term forecasts, do they not?  If you only use the average aging curve of similar players, while it is nice to be able to get some information about how similar type players age, you will get all kinds of freaky results.  You have to use a reasonable blend of the average aging curves of your similar players (where the uncertainty is high) AND the average aging curves of all players (where the uncertainty is low).

I have always assumed that Pecota in general, blends the two methodologies, and I still think they do. But for some of their presentations, perhaps the “percentiles”, I think they rely to heavily on the comps and not enough on ALL players and “theoreticals.” By “theoreticals,” I mean, for example, we know that these error bars cannot be symmetrical, we know that players with less historical playing time will have wider bars, etc.

BTW, as far as the percentiles being symmetrical, don’t forget that a bog component of them is the random variance in X number of PA, and that IS symmetrical.  Combine that with the fact that even the standard error on the true talent estimate tends to me somewhat symmetrical when you incorporate selective sampling and playing time, and you get a total result which should be somewhat symmetrical.

Colin/90: right, and that’s the wrong way to do it.  That’s because the confidence of those comps is far lower for Wieters than it is for Dunn.

It’s one thing to say what Wieters top comps did and what Dunn’s top comps did.  Clearly, we trust Dunn’s top comps more, given that with only 250 (minor league) PA for Wieters, how much are we going to trust his results?

By not putting in any uncertainty around Wieters top comps, Nate is treating his baseline forecast as if there was ZERO uncertainty.  Even if it’s not zero for whatever reason you want to consider, it certainly is not more uncertainty than Dunn’s comps.

***

To convert EqA to wOBA:

wOBA = 1.25 * EqA

(more or less)

So, if someone has a .320 EqA, that would be the equivalent of .400 wOBA.

One SD for .400 wOBA with 600 PA is .022.

If you treated EqA as a straight binomial, you’d get one SD as .019.

On EqA versus wOBA:

I don’t think that works, though. Take a player with a .260 EqA. That’s .17 R/O. A .240 EqA is 0.14 R/O. A .280 EqA is a .21 R/0. In other words, +1 SD is not the same as -1 SD; it’s .04 R/O versus .03 R/O.

This gets more pronounced. A .300 EqA is 0.25 R/O. A .220 EqA is .11 R/O. That’s .08 versus .06.

Your multiplier to convert EqA to wOBA changes as you move up and down the EqA curve - at low values it’s something like 4 times EqA to get wOBA. It then decreases around the average, and then increases again at the high end. (This is holding OBP constant, so that you have the same outs per PA at each level.)

The other issue is that two players with the same wOBA in the same number of PA will have the same value; this is not true of two players with the same EqA. Now I don’t know if that matters for this purpose.

The Wieters comps have a sim index of zero, which means that essentially PECOTA has NO confidence in his comps. So they should definately treat his comps differently than Dunn’s comps. Unlike MGL, though, I am unwilling to assume that they do just because they should; there’s no evidence they do.

(As an aside, I really think that for comps, a big variable should be level. Matt Wieters shouldn’t be being compared to MLB players who are 22 years old, he should be being compared to minor league players who are 22 years old.)

Colin: why are you talking about R/O?  Can you redo your calculations as R/PA?

***

If the Wieters comps shows a sim score of zero (i.e, he has no comps), the obviously his performance expectation has to be whatever the binomial said plus whatever the uncertainty of the mean is.

As I said, 1 SD is roughly 22 points (+/- 22 points).  I don’t know how many SD has 80% of the data points (let’s say it’s 1.5) to give you the 90th and 10th percentiles, so that’s +/- 33, which is a 66 point range.

This is EXACTLY what Nate is showing for all his players (except Pujols).

Nate’s percentiles presumes that the uncertainty of the estimate is zero.

Because EqA = EQR/Out/5^.4

EqA is essentially R/O, except for the weird manipulation to get it on the BA scale. That’s why I did it the way I did. (I do have R/PA estimates, but they’re based upon the assumption of OBP being constant, which isn’t really true.)

You can play around with the whole spreadsheet:

http://www.editgrid.com/user/cwyers/rates_to_r_o

EqA is not R/O, not essentially at all.  It is much closer to being R/PA than anything.  That exponent essentially turns it from R/O to R/PA.

EqA is R/O plus. I don’t think the plus part adds any real value. Yeah, if you want you can treat the part of the curve from about .1 to .3 the way you treat OPS or wOBA or whatnot, as a linear relationship to R/O at about 2:1. But it isn’t, and so it won’t work out exactly that way.

I tested this once already.  If you do 1.8*OBP+SLG, or you use wOBA, you get an almost linear relationship to EqA.

One SD for .400 wOBA with 600 PA is .022.

??

sqr((.4*.6)/600)=.020

Are you treating wOBA as something other than a binomial, which is probably correct?  If yes, how do you get .022?

MGL: I’d recommend reading The Book

To get the binomial using wOBA, you use 1.1-wOBA instead of 1-wOBA for the “q” term, in:
sqrt(p*q/n)

The more HR a player hits (relative to his walks), the higher that 1.1 becomes.  Generally speaking, 1.1 works fine.

Rally:
http://lanaheimangelfan.blogspot.com/2009/04/matt-wieters.html

I looked at the best hitters I could find in the Eastern league who had at least 200 AB there and were age 23 or younger (Wieters was 22). By OPS, they are:

Ron Kittle, 1981 1.119
David Wright, 2004 1.086
Wieters, 2008 1.085
Nick Johnson, 1999
Pat Burrell, 1999
Vladimir Guerrero, 1996
Sean Casey, 1997
Scott Rolen, 1996

...

The other 5 [ed: Note, Kittle was discussed in preceding paragraph] played in the majors the following year, and on average posted an .839 OPS.

I posted this comment on Rally’s blog:

Sean you lost me.  You wrote an OPS for 3 of those players.  Is that their OPS in the Eastern league in the years in question?  Or are those MLE’s? What about the other 5 players?  Why are their OPS’s not listed?

You say the other 5 players (I presume Wright, Burrell, Vlad, Casey and Rolen) posted an .839 OPS the next year in the majors. What about Kittle in 83?  Come one, you can’t just use the .839 one year later and not Kittle 2 years later!  You can age adjust Kittle’s major league OPS by “rolling it back” 1 year, but you can’t ignore it!  Same for Johnson. OK, he was hurt the next year, but what about the year after that?  Same thing, use that and age adjust back one year.

Use all 7 players, give us their major league OPS (either in the next season, or in the case of Kittle and Johnson, the next season after that, adjusted for an extra year of age), and then we can talk.

And also give us the minor league OPS’ for all the players, not just Wright and Kittle!  You say the “best” Eastern Leaguers, but for all we know, those last 5 in your list may have had a minor league OPS substantially less than Wieters!

He noted: “All of these players had an OPS of at least 1.030. “, so I suppose the overall average of the 7 is around 1.065, close enough to Wieters’ 1.085.

I think the point is that the five best of those 7 did an .839.  How you want to handle Kittle and Nick Johnson might skew the results (downward) too much.

I didn’t post the OPS for all players on the list because I moved from one computer to another and had not written them all down.  For the ones not listed, they are in the range of 1.030 to 1.080.

Kittle was in AAA the next year (and it was a monster - 50 homers) and the year after that had an .818 in the majors.  Johnson was hurt the next year, played AAA after that, and a .749 in 2002.

This was a quick look, not some detailed controlled study, so there are no MLE’s or any math besides simple averaging here.  I left out Kittle and Johnson because unlike the others they did not play in the majors the following year, but add them in and the average is .823.  Not that much difference.

OK, got it.  It was just a little confusing.  BTW, I’ll gladly take the over on your projection for Wieters, Rally, and I’ll spot you 10 OPS points!  $100 for charity?  300 PA minimum?

Mike Fast posts in Rally’s blog the OPS of those 7 hitters in each of the next 3 years following their monster AA.  The single highest OPS for any of those players in any of those three years is Vlad at .978.

PECOTA is giving Wieters a mean forecast for his first three years as: .941, .941, .952.

MGL:
And I’ll take the under on your .917 projection, and spot you 10 OPS points.  $100 for charity of your/my choice.  Deal?

Here is a spreadsheet of those players stats.

Guy, my official projection is .870 (I sent my projections to Tango if you want to confirm).  I didn’t tweak that one either.  That is exactly how it came out of the computer using my normal methodology.  The .917 was an off the cuff one, with no adjustments for context (for example, park factors and quality of competition - he is going to be facing well-above average pitching from BOS and NYY)…

MGL:  Fair enough (I certainly take you at your word).  Obviously, that would still be a hell of a year for a young catcher, but not as other-worldly as the Pecota projection.

I think an .870 OPS would translate roughly to a 130 OPS+.  Just for fun, here are the catchers age 25 and under who have done that (1921 to date, min. 300 PA).  Jason Kendall—who knew?

(Formatting advice welcome.)

Player / OPS+ / Age / Year
Johnny Bench 166 24 1972
Carlton Fisk 162 24 1972
Joe Torre 156 25 1966
Mike Piazza 152 24 1993
Rudy York 151 23 1937
Joe Mauer 144 23 2006
Bill Freehan 144 25 1967
Brian McCann 143 22 2006
Ed Bailey 143 25 1956
Ted Simmons 142 25 1975
Johnny Bench 141 22 1970
Tom Haller 141 25 1962
Mike Piazza 140 25 1994
Joe Torre 140 24 1965
Joe Torre 140 23 1964
Rudy York 140 24 1938
Gabby Hartnett 138 23 1924
Joe Mauer 137 25 2008
Gary Carter 137 23 1977
Jason Kendall 136 25 1999
Tim McCarver 136 25 1967
Ed Herrmann 135 23 1970
Yogi Berra 135 25 1950
Brian McCann 134 24 2008
Matt Nokes 133 23 1987
Darrell Porter 133 21 1973
John Orsino 133 25 1963
Jason Kendall 131 24 1998
Ernie Lombardi 130 24 1932

What I normally do is put the numbers on the left, and the name on the right, since the numbers are all the same size.  I’ll edit it for you if you like.

I’ll need more than 10 OPS points, especially since the point of my post is that I’m doubting my projection.  Give me anything up to an .840 OPS, and it’s a deal.

Well, since I have him at .870, I’ll be glad to give you .840 Rally.  I would have doubted my own projection if it came it at .900+, but I don’t have any problem with .870.

It is silly to look for players of the same and position to see how they have done.  That serves little purpose. First of all, a player’s position has almost nothing to do with how they are going to hit other than the position is a proxy for body type.  And catchers tend to get worn out and teams tend to move really good hitting catchers off of the catching position.  But that remains to be seen.  I mean using players at a certain position would be like saying A-Rod could not have possibly hit they way he did because SS just don’t do that.  Yeah, that is because SS used to be and still are to some extent 5-10 175 pounds and wiry.  If you know your SS is 6-3 and 210 pounds you DON’T compare them to other SS!  One reason that catchers don’t hit, besides the fact that they tend to be selected for their defense and intelligence, is that they are traditionally fat and slow and not too tall.  I’ve never seen Wieters and have never seen what height or weight he is, but I am guessing that he is built like a heavy-weight prize fighter, tall and strong.

The other thing is if you are going to look at similar players, you have to at least look at players who had similar minor league stats.  Looking at a list of young catchers without selecting them by their minor league stats makes no sense.  There might never have been a catcher who had an MLE above 1.000 for at least 250 PA for all we know.  So what if I make a list of the best young catchers ever in their rookie years and none of them hit above .820, but none of them ever had an MLE the year before above .900??  What good is that list?

The bottom line is that we have a distribution of true talent among all 23 year old players who are really big and strong (I just looked up his physical size), not catchers.  We have to assume that there is no limit at the tails, otherwise we can never have a player who is the FIRST at anything, right?  So there are players that are a true .800, .700, .820, .900, .950, .600, etc.  OK, so we sampled the guy and he hits at a 1.000+ clip in 250 PA.  If you want to throw in his lower minor stats, that is fine with me.

Now we have all we need to estimate his true talent OPS exactly.  So let’s stop arguing about it!  Isn’t that what all these projection systems are supposed to do?  Take a player’s sample stats, the estimated distribution of true talent for “like players” and then come up with an exact mean of this player’s estimated true talent?  That is all I do.  Anyone who comes up with a number way too low or way too high for Wieters or for any player is doing something wrong.

If Tango, or anyone else, can give us the approximate distribution of true talent OPS for 23 year old players who are around 6-4 and 210 (we can infer that of course from players of other ages - we don’t have to use only 23 year old players to get that data), then the rest is a simple Bayesian problem.  What are the chances that a true .700 player hit 1.039 (or whatever his MLE was) in 250 PA, a true .710 player, a true .720, all the way up to, say, a true 1.000 player.  Then we add everything up, take the proper ratios, and we are done. We have the absolute best estimate of Wieter’s OPS this year.  And whatever you get is going to be a little higher (if you have a min PA to test your forecast), because if he plays at all and does not do well, he’ll get sent back down…

It’s a bet then.  Under .840 I win, over .840 you win.  We can call an exact .840 a push, though that’s unlikely as we can go to more than 3 decimal places.  Under 300 PA is also a push.

“I’ve never seen Wieters and have never seen what height or weight he is, but I am guessing that he is built like a heavy-weight prize fighter, tall and strong.”

You guess right, he’s 6’5.

It’s a bet!

Yeah, I looked him up after I wrote that.  I just figured that for anyone to rake like that, he must be big and strong.  That makes a big difference in terms of regressing someone with few PA. I mean, if I told you that a 6-5 inch 230 pounder, had a .900 OPS in 100 PA and another guy, Pedroia-like, had the same stats, what do you think the difference in projections would be? A lot!

MGL:  I don’t know if your #114 was a response to me, but I said nothing about the relevance of position for making projections.  I just thought it was interesting to see the company Wieters would be in if he hits his projection.

But since you raised it, I’m not sure you’re right that it’s “silly” to consider position in this context.  And your statement that “a player’s position has almost nothing to do with how they are going to hit other than the position is a proxy for body type” is backwards.  Catchers and SSs hit worse because they’ve been selected in large part based on defensive ability—the body types are a residue of that selection, not the other way around.  Catchers are worse hitters simply because they can catch, and it is less likely for a human being to have two exceptional skills than just one.

So perhaps position should be part of our prior estimate for Wieters.  We’re basically asking “how likely is it that someone is truly an .870 OPS hitter.” If that probability is X, then the probability that someone is an .870 hitter AND can play catcher at the MLB level must be something like X/5 or X/10.  (There have been 119 players since 1921 with a career OPS >130 (3000+ PA), and 2 of them have been catchers.)

I certainly agree that height/weight should be considered as well.  But I’m not sure it totally overrides position as a consideration.  Suppose all we knew about two players was they were both 6-3, 220 lbs., and had an OPS of .950 in 250 PAs in the same AA league.  One is a catcher and one is a first baseman.  You’re saying it’s a coin toss as to which player hits better next year; I think I’d put my money on the first baseman.  But I’m not sure…

I agree with Guy.  I don’t use position to adjust projections, except in the case of catcher.  Shortstops may hit less on average, but I agree that if you have a shortstop built like a RF, who hits like a RF, then don’t hold SS against him. 

But catcher is different, because they will take a beating.  And if you move him off of catcher, he will hit better.  I’d be a lot more scared of losing the bet if the Orioles announced today that Wieters would be called up today, but will play 1B or DH.

It won’t happen though as Wieters is also a good defensive catcher.

Right, catcher is an exception because of the beating they take.

***

There are two issues:
1. the actual wear and tear of a position
2. the selection bias of the player

ARod, for example, would regress to the same mean, regardless of the seven fielding positions he was in. 

Frank Thomas would regress to a different mean.  If for example his hitting was scouted as being so lights out that no matter how bad he was as a fielder that a team will find some place to put him on the field, then his population mean would be even better than the average 1B to regress toward.

ARod, for example, can be scouted similar as Frank Thomas, but because he’s also a good fielder, it would mean that he might regress to a lower mean.  Again, that makes no sense.  Unless we “know” that scouts love ARod’s hitting as much as they loved Frank Thomas.  It’s based on what can we infer.

The problem is that it’s tougher to figure this out for the more gray-area players, and so, we infer scouting based on their fielding positions.

***

Catchers have an additional consideration, that they may lose say 10 runs per 700 PA because they catch.  So, Wieters would need to regress to a lower mean because of that.

Catchers are worse hitters simply because they can catch, and it is less likely for a human being to have two exceptional skills than just one.

Guy - This seems to me to be a questionable statement.  Catching and hitting both have learned components that are specific to their skillset.  But they are also based an a lot of shared genetically determined abilities like good eyesight, quick reaction time, strong legs and arms, and intelligence.  I think that the only reason that you don’t see more catchers that are good hitters is that good hitters have more options about where they can play when they are young and since few young players actually want to catch the good hitters self select other postions.

Peter:  The relationship between defensive demands of a position and offensive production couldn’t be clearer.  The more important defensive skill is, the smaller the resulting talent pool must be, and thus the lower average offensive production we expect (i.e. the fewer excellent hitters available).  And that’s exactly the pattern we observe.

If a lot of good young hitters had the skills to be MLB catchers, their coaches/agents/fathers would have every incentive to push them to catch.  Because their chance of having a successful professional career are greatly magnified if they can catch (replacement level offense is set much lower).  So I find your theory totally implausible as a generality. (However, once a catcher reaches the majors, he may have an incentive to learn another position and try to extend longevity.)

I don’t doubt that in the general male population, hitting and fielding talent are positively correlated.  But at the level of MLB, they will always be negatively correlated.  (The extreme examples of course being pitchers, who are selected 100% on their defensive talent, and DHs on the other side.)

BTW, here’s the breakdown of hitters with OPS+ of 130 or better (3000 PA), by position:

C 2
SS 2
2B 3
3B 7
1B 26
OF 69

Let’s say you have a really good hitter, +30 runs per season, or +50 vs replacement level.  He’s an average defender at catcher, which is pretty good considering the high level of defensive talent among MLB catchers.  As a 1B, he’s also average, he’s slow and the skills he has at C don’t translate anywhere else.

His value, with position adjustment (+10 C, -10 1B)

1B +40
C +60

Except that if he catches, he loses 10 runs on offense because of the beatings he takes, and instead of playing 150 games per year, he only plays 120.  You have a replacement level catcher for the other 30 games.  That makes the comparison:

1B +40
C +40 (50 +10 pos adju, -10 off adj)*120/150

So if a guy is below average defensively at catcher but a good hitter, he might have more value at a position where offensive value is easier to find.  A lot of that value comes from being able to stay in the lineup more often.

Makes me wonder if Mike Piazza would have been more valuable had he been moved earlier in his career, like Carlos Delgado was.

Or if Kendry Morales doesn’t start out hot, and Mike Napoli’s shoulder keeps hurting....

Rally:  If we assume a hitter loses 10 runs just from playing catcher, doesn’t our position adjustment then need to me more like +20 than +10?  Otherwise we’re saying that finding an average catcher is as easy as fining an average 3B.  Do we think that’s true? 

Your point also makes me wonder if good hitting catchers shouldn’t migrate to the AL, and then DH on their off dates behind the plate.

Position adjustment is still +10.  The beating they take and resulting loss of offense is the position adjustment.

I don’t have my THT in front of me, but I thought Tango found some evidence that catching was holding back offensive ability.

Rally:  I don’t care if you include the 10 run “catcher penalty” in the position adjustment, or lower replacement-level offense at catcher by 10 runs (so now he’s +70 before the -10 adjustment), but you need to account for it somewhere.  The team will pay that 10 run penalty no matter what, so it’s just a constant:  your catcher will be 10 runs less productive there, but so will a replacement catcher (or anyone else). 

The only reason it would matter is if we think the effect is multiplicative rather than additive, i.e. catching reduces offense by 12%.  In that case, you wouldn’t want to put a great hitter there if you could avoid it.

BP actually has an article on catcher hitting today:

http://baseballprospectus.com/article.php?articleid=8706

I don’t subscribe, so I can’t say if it’s any good or not.

I agree with Guy.

I don’t use position to adjust projections, except in the case of catcher.  Shortstops may hit less on average, but I agree that if you have a shortstop built like a RF, who hits like a RF, then don’t hold SS against him.

I guess you don’t agree with Guy.

Guy says that you should hold it against him - that the SS built exactly like the RF will regress to a different mean.  You apparently disagree, as does Tango.

I was going to say that Guy had a point, which is a mathematical/Bayesian one, but since Tango and Rally do not agree, I think Guy may be wrong, but I am not sure yet.

Let’s take the SS which is a better example.  We have a 6-3, 210 pound SS, like an A-Rod. He hits a lot in a sample of hitting performance. The question we ask is that for our Bayesian “priors” , do we use the population of all 6-3, 210 pound hitters period, or do we somehow temper that because he also has the skills to be a SS.

OK, I think the answer is that we use for our priors a 6-3, 210 pound player who also is agile and quick enough to play SS, which actually might increase the numbers.  The fact that there are so few people in the world that are that big AND possess the skills to play major league SS don’t enter into the equation, I don’t think.  But again, I am not sure.

I just posted that, and I think that Guy may be right (and Tango and Rally wrong)!

What if it is exceedingly rare to have a player who has great hitting skills and the ability to play SS?  So rare that none have existed in the history of baseball.

So now, we have a SS, like A-Rod, who has a monster sample hitting performance and we are trying to determine his true hitting talent.  If we use the priors from all big players we are going to be using lots of players who have true talent hitting that are great. But, I just said that there is no such thing as a player with SS skills and one that hits that well.  So, we have to adjust those priors, no?  We can’t use all those players at other positions who are great hitters in our priors, can we, if we just said that it is virtually impossible for a player to have SS skills and be a great hitter. Again, I am not sure.  The player in question (A-Rod) may be the first and we don’t want to eliminate that just because it has never happened before.  But what if there is actually something about the skills necessary to play SS that makes a person, even a large one, not likely to be a great hitter?  I think that is a different story, which I don’t think Guy is arguing.  He is simply arguing about the chances of a player being a great hitter and having the skills to play SS, which is admittedly quite low. But that may not go into the Bayesian equation.  OK, I am not sure again.  Greater minds than I need to answer this one.  Maybe it is a good question for Andy, wherever he is.  Or maybe someone just needs to put it down on paper, which is what I sometimes do for these Bayesian things…

Guy,

Let me clarify.  Same hitter, but hitting differently depending on where he plays.

For both, I’m adding in +20 per 150 games for replacement level, but the 1B gets a -10 adj and the catcher a +10 position adj.

pos LW vsrep per 150
1b +30 +40
c +20 +50 (+40 in 120 games)

The replacement level catcher who plays the other 30 (well, really 42) games is a -30 hitter per 150 and average defender.  If he were not catching, he’d be a -20 hitter.

All numbers for theoretical discussion only.

I don’t see how we can reduce our estimate of a hitter’s true talent (ARod) the better the fielder he is.

If it’s true that our prior is lower in such cases, that simply means we are not considering enough information.

There is an inverse relationship between the quality of hitter and quality of fielder.  This is a fact, since the average at each position is (roughly) league average, and since the worst fielders are at LF/RF/1B is also where we find the best hitters, we know this is true.

If all we knew was this, then we know that ARod is drawn from the population of “infielders”, then you regress to what you know.

But, we know more about ARod.  We know that he is a good enough hitter that he could play any position.  We know this not because of his sample data, but because he likely would have been the 1st nonpitcher selected even if he was a 1B. (Derek Lee was the first 1B selected that year.  Trot Nixon was the first nonpitcher selected after ARod.)
http://www.baseball-reference.com/draft/?year_ID=1993&round=1

So, it is an insane proposition to want to regress ARod to a different mean than Frank Thomas or Albert Pujols (is he SS, 3B, LF, or 1B… depends when!).

All this tells me is that, yes, most of the time, it works out that way.  But, if you know something specific about someone, and if you are going to evaluate someone specific, then you have to include the additional criteria that makes you talk about that someone.

If we just talked in general terms, we’d be ok.  It’s because we talk specifically about ARod and Wieters, then we need more unique identifying characteristics.

Tango:  I agree it feels wrong to regress A-Rod to a lower mean.  And certainly we would take account of his size.  But let’s say accounting for his size, we regress him toward a 3B mean, while regressing Thomas to a 1B mean.  (Realistically, it won’t change our estimate a lot, but it’s an interesting puzzle.) In fact that would work out fine as ARod was not quite the hitter in his 20s that Thomas was. 

The key is that we aren’t just learning that ARod “is a better fielder” and therefore lowering his forecast.  We’re learning there is a reason other than his bat that he is in professional baseball.  That’s not true of Thomas.  So we know that one guy is playing only because he can hit, while the other need not be as good a hitter to be where he is.  Doesn’t that change the likelihood that each is really a .900 OPS hitter?

In real life, a couple of seasons of real performance data, plus scouting info, would probably swamp the effect we’re discussing.  But I still think (pending Andy setting us all straight), that I’d bet on a first baseman with Wieter’s identical stats to outperform him.

* *

Rally:  Maybe we’re talking past each other, but I don’t think you’re factoring in the effect of the catcher penalty on replacement level.
At 1B, your player is +30 LW, +50 vs repl, -10 for position = +40. 
At C, he is +20 LW, which is +50 vs. a replacement player who also suffers the catcher penalty, +10 for position (defense) = +60.  (+48 in 120 G)
Bottom line:  the 10 run catcher penalty is a wash—it should never affect our decision on where this guy should play.  If he doesn’t lose the 10 runs, someone else will.

Right.  There’s only one reason Frank Thomas was drafted.

But, if we KNEW that ARod would STILL have been the overall #1 nonpitcher drafted, if he had Thomas’ glove, then we need to know that, right?  Once we know that, he gets regressed to the same mean as Frank Thomas.

Looking at it the other way - it should matter if we want a true talent estimate or a forecast, right?

In other words, for a 1B/DH/corner OF type, if we want true talent, we should regress his stats to the population mean - with the caveat that we can get more selective about his population.

But if we want to have the most “accurate” forecast, we should definately regress to the positional means. If we have a 1B/DH type who projects to a .680 OPS or so, then we don’t expect him to play much - he’s going to play himself out of a job.

But if he plays above his projection, he will get playing time and thus enter into our evaluation sample. So if we regress based upon position, we’ll get a more “accurate” forecast even if we have a worse estimate of true talent.

But, if we KNEW that ARod would STILL have been the overall #1 nonpitcher drafted, if he had Thomas’ glove, then we need to know that, right?

We obviously don’t know that, Tango, whether A-Rod would or would not have been a first round pick had he been a DH in college!  You can’t assume that after the fact because it turned out that he was such a great hitter! So, yes, I think Guy is right.  We know only 3 things generally (if we did more research, we could find more of course).  A player’s physical data, his position, and his performance data, both offensive, defense and otherwise. The question is what to regress to, given that.

I think that Guy is right because we are less certain that he is a great hitter because he COULD have been drafted as a not so great hitter because of his fielding skills.  If he were a DH in college, then we know that he was drafted only on his hitting skills. So a player’s position, at least in the past and maybe mostly in amateur and minor league ball is a clue, albeit a small and roundabout one, as to his true hitting skills.

Colin that is true to some degree, and hence, the difference between what I have been talking about lately, a “forecast” (that assumes a certain min level of playing time) and an estimate of true talent, that involves any amount of playing time, although even then, you can’t quite test the accuracy of the latter because of the possibility of no playing time.

So basically, I think you regress to a mean that is some combination of a player’s physical traits (and age of course, although that is also related to physical size - the two are not independent) and any other “clues” you have as to his hitting talent, somewhat independent of his stats, such as his defensive position.

So much easier to just use similar players, which incorporates everything automatically.  Enter Pecota…

Guy, in my example the catcher is +20 runs vs average, and +40 vs replacement level.  In your example it looks like you are making the difference between average and replacement 30 runs at catcher and then giving a 10 run position adjustment on top of that.

I see the replacement level players as follows:

1B -10 runs offense, avg defense
C -30 runs offense, avg defense.

If that replacement level catcher was moved to first, he’d be -20 runs on offense, therefore below replacement level overall for that position.

At the same time I’d guess the 1B would drop to -20 runs if he had to catch, but probably also lose 20-30 runs defensively.  He’d be way below replacement level as a catcher, because he doesn’t have the skillset and experience.

You can’t assume that after the fact because it turned out that he was such a great hitter! So, yes, I think Guy is right.  We know only 3 things generally (if we did more research, we could find more of course).  A player’s physical data, his position, and his performance data, both offensive, defense and otherwise. The question is what to regress to, given that.

I’m not assuming after the fact.

And if all you have is what you said, then yes, we don’t treat ARod any differently. 

However, for 1st round picks we know alot more than just that.  And for ARod in particular we know alot more. 

So, to put in a one-size-fits-all type of regression that ignores the likelihood that ARod was considered a better hitter than Derek Lee when they were drafted, doesn’t mean we get to still regress Lee to a higher mean.

If we talk about specific players, then we need to talk about additional characteristics.

If we talk about general terms, then we treat ARod like a SS and regress him to a lower mean.

Colin #98: “EqA is R/O plus.”

What does he mean by “R/O plus”?

terpsfan -

The equation for EqA is:

(EqR/O/5)^.4

Thus, EqA is R/O translated onto the batting average scale. Tom is correct that, in the main part of the curve, EqA roughly correlates to OPS/wOBA/etc. - you see the same general 2:1 relationship.

But it’s not precise - there’s a bit more of a curve to it. And you do end up with the problem that two players with the same EqA and the same number of PAs are not necessarily equally valuable as hitters.

The main point I had was that since EqA and wOBA have a fairly strong linear relationship (at the level of MLB players), and since we’ve shown in The Book how wOBA can be used in an almost binomial fashion, the same holds for EqA.

So, if you want to know how often a true .300 EqA hitter will hit at under .280 or over .320 in 600 PA, we know almost exactly the answer.

By plus, I thought you meant EqA accounted for the extra runs created by a player’s teammates based on the player’s rate of avoiding outs. EqA does not account for indirect runs created.

It’s getting close to collection time on the bet detailed in posts 105-116.  Matt Wieters has struggled enough that the only doubt I’ve had for a while is the chance of an injury stopping him before 300 PA.  He passed that mark last night.

I’ll probably pick retrosheet but am open to suggestions.  Given some of the recent debates I might ask for the charitable donation to go to a disease-fighting cause, some form of cancer or diabetes.

Of course, the year isn’t over yet.  Wieters could salvage MGL’s bet by hitting 15 straight homers or something.

I’d like to see the Wieters forecast on the back cover of BPro 2010.

***

As for charities that are baseball-related, you can consider MGL donating to a little league in Rally’s hometown (in Rally’s name of course).  I know that my kid’s uniforms are sponsored by some financial firm or the local pizza place.

I’d like to see BaseballProjection.com on some kid’s uniform!

I don’t have any pull there, Tom, but I’ll see if I can talk Studes into putting it on the front cover of the THT annual.

(To be honest, I’m not even so interested in them fessing up to what happened with Wieters. But I really would like them to at least tell us they’ve identified and fixed the problem - the DT-based MLEs on their website are so tempting sometimes, and it would be nice to know if they actually WORKED or not.)

Just let me know where, how, and when, and I am happy to oblige!

It’s funny.  I mentioned this before but I am not a big fan of the AAAA concept (and I am not suggesting that Wieters is a AAAA player, and I don’t think the scouts think so either), but the first game I saw him bat, I thought, “Crap, this guy looks like he is going to have trouble in the majors.”

I’ve never seen him in the minors, but I have a hard time picturing him tearing up any level of baseball.  He just does not have a good approach at the plate, at least at the major league level.  Maybe there is something to the AAAA concept. If there is, shouldn’t scouts have been able to see it in Wieters?  I noticed it in his first major league game (of course, easy to say after the fact).  Shouldn’t they have been able to see something in the minors or in his swing or his approach (or mental makeup) that would have suggested he was not going to be so successful in the majors, at least early on? Maybe they did and we just don’t know it.  After all, the scouts don’t do MLE’s, I don’t think.

Or maybe he was very lucky in the minors.  That is always possible you know.  There is likely always some degree of luck in any above-average performance over any finite time period. We NEVER know how much luck there is. It could be from negative infinity to positive infinity (IOW, it is possible, from the numbers alone, that Wieters is actually the worst or the greatest 6 foot 4 inch catching prospect that has ever graced a baseball field)…

I’ve never tried to project from college stats but Brian C’s stuff suggests that Wieters was kind of lucky in 2008.  I think he’ll wind up hitting a lot better than he’s done in 2009.  My guess is I’ll still project a slugging % in the .425-.450 range for him.

In his rookie year I’ve seen a player with terrible strike zone judgement.  This is a bit unexpected given his track record.  My observation is his biggest problem is not swinging at bad pitches (though he’s done enough of that) but taking too many hittable pitches.

I think one lesson here is that AA is a lot further away from major league quality than some people were assuming last winter.  The Eastern league is not very close to the majors, and it does not have a higher talent level than AAA.

I still don’t believe in the AAAA concept.  It applies to about one guy, Brad Komminsk.  Maybe David McCarty.  There are a lot more guys, probably hundreds, who might have been written off as AAAA and eventually prove to be able to hit MLB pitching to some extent.  Guys like Randy Ruiz or Oscar Salazar.  Of course Ruiz only got a chance to play when Toronto finally lost patience with a MLB veteran, Kevin Millar, who a decade ago was another of these “AAAA” players waiting for a chance.

Come to think of it, I don’t think one directly replaced the other, but Kevin Millar on the Red Sox took the role that another former AAAA player, Brian Daubach, had before he outlived his usefulness.

Clearly there are players that have great MLE’s for a long time, but end up performing badly in the majors for a while and then get dumped, sometimes for good.  But, since that will occur (a lot) even if there is no such thing as a AAAA player, that is not evidence that the AAAA concept exists…

So Matt Wieters finishes with a .753 OPS.  The Orioles can be happy with his strong finish, an .882 OPS in September/October.  And since I resurrected the thread on 9/15, he hit 373/422/587.  Maybe MGL just needed a little more season for his forcast to beat mine.  We’ll find out next year.

I request a charitable donation to be made to the Baltimore Humane Society.
http://www.baltimorehumane.org/

Last year in the Rays/Red Sox playoffs, I was the first person to take one of MGL’s bets, at least in a baseball related forum.  Now, I think I’m the first to win one, I think.  Anyway, it was good fun (both winning and losing).  Thanks MGL.

Bumping this thread up for MGL.