THE BOOK--Playing The Percentages In Baseball

What do you think the starter/relief split should be?  Relievers this year are pitching 35% of the innings.  They probably should still have a smaller % of WAR than that, since the average reliever is closer to relief replacement level than starters are, but the fact that the best relievers get the high leverage should offset that somewhat.

I chose a higher replacement level for pitchers because it didn’t look right when I applied it to actual examples.  In a 4.50 ERA league, a .420 pitcher has a 5.40 ERA, a .380 pitcher has a 5.90.

From running the projections, it just seemed there was a huge number of pitchers who can’t find jobs, are stuck with minor league contracts, or are constantly on the waiver wire who are better than .380 pitchers.

Rally, for pitchers, I use .380 / .470 as the replacement level for starters / relievers.  The gap must be somewhere around a .090 wins per 9IP difference, since this translates to around a .90 ERA difference of 1.00 RA difference, which is what we think happens to pitchers who switch between starter and relief.

The average starter is around a .490 pitcher and the average reliever is around a .520 pitcher.  We can further break down the relievers as (purely as an educated guess):
one-half .480 pitchers (with LI of 0.7)
one-third .530 pitchers (LI of 1.1)
one-sixth .620 pitchers (LI of 1.7)

Let’s give starters 65% of the innings.

Let’s work it out.  .490 minus .380 is .110 wins per 9 innings.  Multiply by .65 and you get .0715 WAR per team game played.

Now for relievers.  Remember that we are giving an “effective” LI of LI/2+.5.  So, we get:
(.480 - .470) * 0.85 * .5
(.530 - .470) * 1.05 * .333
(.620 - .470) * 1.35 * .167

We add it up, and we have +0.059 WAR per 9 IP.  Multiply by 0.35 and we get .02065 WAR per team game played.

So, per 162 games, we have:
11.6 WAR, starters
3.35 WAR, relievers

We are speaking the same language on relievers, since you have 3.4 per team. Can’t get any closer, especially since I had to estimate my reliever model for our purposes here.

For the starters though, I can’t get behind yours.

I projected 728 starting pitchers last winter, pitchers from the big leagues, AAA, AA, and A+.

By my level, 406 of them rate as at least 1 run above replacement.  Yours is about a 0.50 difference in ERA, so you’d have by that measure 533 above replacement starters.

If anything, it makes my wonder if my level is still too low.

Pitchers are so hard to project anyway, but I looked at some faces that I had projected at around .420:

Jeff Suppan
Livan Hernandez
Victor Zambrano
Adam Eaton
Tom Glavine

Suppan hasn’t been that bad, and makes a lot of money, Livan got more than the minimum and has been pretty solid for the Mets.  The other 3 have been replaced.

Some projected at .380:

Steve Trachsel
Mark Mulder
Rick Porcello
Sean O’Sullivan

Two who couldn’t get jobs, one outstanding young prospect who the projection system frankly whiffed on, and a guy with a 6.06 ERA in AAA who just gave the Angels a 7 inning, 1 run win in his debut.  A lot of minor leaguers who I know little about around both the .420 and .380 level.

But what really gets me is the guys I have projected above .420 - at 3 to 5 runs above my high replacement level.

Kenny Rogers
Phil Dumatrait
Jeff Marquez
Matt Chico
Chase Wright
Sean Smith (damn, a replacement level even I surpass?)
Sidney Ponson
Shane Loux
Giancarlo Alvarado
Jason Windsor
Runelvys Hernandez
Kris Bensen
Lenny DiNardo

and many more.  I have Odalis Perez, who nobody would give a guaranteed contract to, at +10 runs above replacement.

So I’m pretty convinced .380 is way too low.

If we accept that .420 is what a replacement level pitcher would do as a starter, how do you think that same pitcher would do as a reliever?  I can accept .420 for a SP if you go at least .500 for a RP.  This is what MGL (effectively) uses I think.

What is clear to me is that there is a TON of starting pitchers who could be half-decent relievers, but choose not to, while there is a TON of pitchers who want to do ANYTHING, and will relieve because there is a spot there.

As a way to figure out whether the 70/30 split you are using or the 78/22 split I am using is better, we can figure out how much SP get paid (above the minimum) and how much RP get paid (above the minimum).  I figure it’s at least 75% for SP.

I split pitchers based on their ages (25 and under, 26-28, 29 and over).

If I did this correctly, I get that 81% of the 25 and under salaries (above the minimum) goes to starters, 88% of the 26-28 pitchers goes to starters, and only 67% goes to starters if they are 29 and over.  (Teams love them some old reliable relievers.)

The percentage of innings for the three classes is 77% of 25 and under are starters, 76% for the 26-28 pitchers, and 64% for the 29 and over pitchers.

Data is for 2004-2008.  Numbers look a bit off.

Basically, it looks based on this that Rally is more justified in using 70/30 as the split, even though we clearly have a bias here in terms of the old pitchers.

I’d have to check what baseline I’m using for relievers, but I think it is close to .500.

That’s about what I said on my projection explanations (4.55 ERA for 2009), but who knows what I put into the WAR database.

I’ve got starters set at .420 replacement level and relievers at .480.  I probably should up the reliever level to about .500, next time I rebuild the pitcher database.

Anyone who’s already bought my data don’t worry, If I make a major change like that on the site I’ll provide you an update free of charge on the download.

The free update is nice to be sure, but I was kind of hoping there’d be no major changes going forward.

I’m not sure how many people realize it, but Clay tinkered with his WARP system quite a bit from year to year and I always found that frustrating as my numbers would all of a sudden not match the newly published ones.  I recall one year that the catcher data all changed a fair amount - he must have made some specific adjustment to C baseline or defense or something.

If you changed the releiver baseline would that just change the releiver data or all players as the relative ratios would have to be changed as well?

Just in terms of the starter/releiver split - and yes I realize this is faint praise - but WAR is much more reasonable than the old WARP data that I’m replacing.  There are lots of examples of long career releivers who ranked ahead of medium length career starters in old WARP, but in WAR the starters (appropriately) rank higher. 

I have been a bit surprised at how highly some releivers rank within their draft classes, which I guess could stem from a too low baseline.  But mostly these are just very good releivers I think.  A career releiver ends up as one of the 5 most productive pitchers more often than people may realize.

I wanted to quickly look at the Pitcher/Hitter split from the drafts that I have finished (1987-1994).

Tango found a 57.7/42.3 split from the 1999-2008 teams.

From these draft classes the split is 67/33 - a much more pronounced split in favor of the hitters.  I’m not quite sure what to make of that.

It does seem to have moderated a bit.  In 1987-1989 the split was 71/29.  In 1990-1994 the split was 64/36 (still higher than what Tango found).

Anyway, I just wanted to let you know, Rally, that I’ve found the WAR data I bought very useful and fun to play around with.  Thanks.

The difference is that Rally is making it easy for you to reupdate your data.

In a database, you have your data, and then you have your queries that run off that data.  You should be able to drop and recreate the data, and the underlying queries need NO changes at all.  If that’s how you are setup, then you are good.  If not, then you are doing too much manual shenanigans.

I’d be happy to look at your model and suggest improvements if needed.

If not, then you are doing too much manual shenanigans.

Guilty as charged!

It’s a real low tech and manuel operation that I run unfortunately.  It has been my intention to get up to speed on things like querying a simple database and I’ve been using this changeover to WAR as a catalyst to that process.

I’ve never had any kind of IDs (aside from player name) in any of my tables.  So one of the big things that I’m doing now is adding in Rally’s retroIDs.  Correct me if I’m wrong, but it’s my understanding that with those IDs in place I will then be able to do mapping and querying that will make updates more dynamic and easier.  Is that correct?

And although I will freely admit that what I’m about to say includes a lot of rationalization of a time consuming and inefficient process… I do beleive that there is some benefit to manuel, labor intensive data entry and manipulation in terms of immersing oneself in data.  The time spent ruminating on things as you slowly build up your data gives you time to think about connections and patterns in the data.

When Rany started his big draft study he made a big thank you to an intern who I assume did most of the data entry.  I know he missed some things because he was not immersed in the data.

But yeah, freeing up time to do other things is a huge benefit too.  I’m trying to get there as well.

The immersing of the data and getting your hands dirty is good.

What is not good is repeating steps.  Anything that can be done iteratively is the reason for the existence of computers.

In MS Access, it even gives you the option to directly link to a text file.  I presume Excel also does that.  I can give you a step-by-step on Monday for that.

For example, I have batting, pitching and player data from several sources (college, minor, major, Japan, Korea) that I want to merge into a single database. The originals are in different formats, so I have to rearrange, link, etc to get the data into the combined tables.

The important thing is to do all the conversion and merging with queries, so that if I ever have to replace or update one of the source tables, I only have to rerun the query.

I hear you Philly, and I’ll try not to make constant changes.  But let’s face it, What’s up there is not perfect and never will be, and if improvements are identified, I’d rather have better data on the site than just be stuck with what I first publish. 

There will be fairly major changes this summer, and I’ll make minor changes (like reliever Rlevel) at the same time as I do the big things (add crude WAR data for players back to 1871.)

What is not good is repeating steps.  Anything that can be done iteratively is the reason for the existence of computers.

You’ve got me there for sure.  But one thing I’ve tried to do is every time I do a full scale redo I add more data so that the database isn’t just updated it’s significantly improved.  I think the first big draft study I did was in 2000/01 and I probably only had 5 or 6 columns of data per player.  The new one will be more than double that.  This year in addition to adding in retroIDs and changing from WARP to WAR I’m including PA and IP which was a suggestion that Tango had made to me a couple years ago. 

I haven’t really looked at those data systematically, but there has been some interesting little tidbits that have popped out.  I did the 1994 draft recently and noticed that the ever fragile Carl Pavano has the 3rd most IP from his draft class with a good shot at finishing 2nd.  Fragile is sometimes about context I guess.

Rally

I certainly agree with you that when you think you have improvements to make you should.  It would be great if you could communicate the timing of those changes and also try to lump in multiple changes in one fell swoop.

What always bugged me about Clay is that he did constant minor tweaks without any public notification and (until this year) left the biggest mistake (his replacement level) in place.  I mean, I guess, it’s nice that he cared enough about his stuff to be doing minor tweaks here and there, but doing them on top of a huge mistake that just sat there always seemed a little crazy to me.

Has Clay put the new WARP in place yet?  I looked at Clemens (175 WAR) and it’s high.  175 WAR on 4917 IP (or 546 9-inning games) means +.320 wins above replacement per 9-inning game.  Clemens is roughly a .667 pitcher, making replacement level as .667-.320 = .347.

I think Clay has put in the new WARP (checking now, Bonds is at 190 WARP for his career, while I remember he used to be more like 230,) but his replacement level is still lower than the one used by you and MGL (and Rally and Fangraphs and Dan Rosenheck...)

Let me email him…

I’d be curious to hear what he says.

Naturally, I’ve noticed some big changes going from old WARP to WAR.  I haven’t done a lot of spot checking against the new WARP, but there are still some differences in high PA replacement level players.

The idea of how much playing time a sub-replacement player would actually get came up in another thread here so I’ve been paying attention to high PA players with negative WAR.

Just from the 1988 draft:

Mark Lewis: 3007 PA with a WAR of -3.9.  His old WARP was 8.5 and his new WARP is 3.0.

Ricky Gutierrez: 4044 PA with a WAR of -3.8.  His old WARP was 20.5 (!) and his new WARP is 4.0.

Rico Brogna: 3192 PA with a WAR of -1.0.  His old WARP was 23.3 (!) and his new WARP is 3.9.

Seems like WARP still has a hard time pushing high PA utility infielder types below replacement level.  I’ve also noticed it with longtime backup catchers too.

PS Rally - your data has Brogna with 3192 PA and BP and b-ref both have him at 3223.  Any idea what happened there?

Yeah, I’m only using AB+W+HBP as plate appearances.  I suppose I could add in sacs, but it doesn’t affect the ratings and I’m not claiming to be a site for official baseball totals.  Sean Foreman has that covered, which is why I link to his site instead of reproducing every batting line along with my ratings.

The reason is that SF has null values for some of the years, and I did most of the work before finding the nz function in MS Access.

(turns null values into zeros for those not familiar with Access).

Rally, since you are using Retrosheet, why not use

(Note the minus sign.)

This way, it matches to exactly the data you have.

I’m using baseball databank for the batting numbers.  You can’t get a complete batting line out of the earlier retrosheet files, there are some missing games.  Otherwise it would be no problem, retrosheet doesn’t have null values.