THE BOOK--Playing The Percentages In Baseball

Tom,
I’d love to see you tackle my question from the mailbag:

Scientifically exploring optimal 25-man roster construction for both an AL team and an NL team.

I think you could get a good, hefty chapter worth of talking points there.

Starting with an argument for carrying 11 vs. 12 pitchers in NL vs. AL; the quantified value of an elite pinch-runner; the importance of 2 lefty relief specialists at least (or, how many lefty relievers is excessive - 4?); platooning and how to best utilize it…

At least 4 or 5 items there that can be tackled sabermetrically, I think.

What I’m struggling with is how you’d throw in hypothetical Win Share values for each of the 25 guys to make it more tangible for readers. Maybe you could just dissect the actual 25-man roster construction of an ML team in ‘08, and discuss whether they optimized their roster.

(Example: The A’s carried Rajai Davis all year since acquiring him off waivers, despite the fact that he’s a poor hitter.  But have his defensive contributions in CF and pinch-running for Cust/Frank Thomas in fact made him the ideal 25th man - a guy you don’t need to give PAs to, but that can contribute greatly in high-leverage situations?)

Thanks for always providing thought-provoking topics and dialogue. The blog is required daily reading for me.

One thing I’ve always enjoyed is analysis of individual player skills.  Tom, I think you liked my piece comparing Kenny Lofton and Steve Finley’s range.  I’m sure there are other interesting side-by-side pieces - a look through the PBP to see how fielders interact, for example? 

- Fielder aging curves.  When do fielders reach their peak level of performance?  Not sure how to do it given rightward shifts on the defensive spectrum.

- Does shifting rightward on the defensive spectrum decrease a player’s speed faster than if he stayed at a tougher position?  Bill James once postulated this about Gregg Jefferies moving to 1B.

- High school baseball analysis.  There might be enough data for the top California/Texas/Florida leagues that you can see what the levels of competition are...Create MLEs...See what batter-pitcher matchups look like for major-leaguers when they were 18.

- Is there really such a thing as a AAAA hitter?  When 28+ year-old career minor leaguers are given a chance in MLB, what’s the range of performance?  From the PBP, do you see they get eaten up by real major-league pitching?

Hah.  I just noticed your ‘Fielding Aging Curves’ at the top of the page...I’ve even read those pieces before...Not sure why I forgot…

1.  Aging curve analysis for a broader variety of offensive skills including contact rate, plate discipline, and power - in so far as certain skills can be isolated from each other.  HR totals are affected by both contact rate and power.  But ISO presumably has less correlation.

2.  The creation of a set of player archetypes capturing as many players as possible without going beyond roughly 10-12 types.  This could be built around something like the classic five “tools” using a more sabermetric definition for each.

We’ve heard/read a lot comments like “guys like so-and-so” age well.  So how does each archetype tend to age?

2.  Same thing but with pitcher repitoires instead, perhaps including other descriptives including handedness, body type, and endurance(?) which may have a larger role for pitchers.

Rick:

#1 http://tangotiger.net/agepatterns.txt

As for #2, that was actually kindof the article I was starting, but stopped, figuring it was too “boring” for an annual.

I want to see all the swinging strike percentages, called strike %, etc used to create a predictive metric.  These numbers should be able to be used in a way to eliminate even more luck than a metric like xFIP does.  It would be interesting to examine.

It would be interesting to look at how the player talent distribution has changed over the years.  Can we extract things and say that offensive shortstops are a higher team priority (at the cost of defense) or is it just that we have more natural talent at the shortstop position now.  Where are the best players playing these days?  Which positions over the years have seen talent gluts or scarcities?

I tried to do the following idea originally a little while back and just didn’t have much luck.  OPS+ normalizes a player’s OPS for that year’s average OPS.  Comparing two different players in two different decades and trying to extract which player was better using OPS+ is not as trivial as comparing their OPS+ if the standard deviation of OPS in the league has changed.  As baseball has become more competitive over time, the standard deviation of the league should decrease since all players are elite and less filler exists.  It might be better then to normalize OPS+ by the variance of a particular year to turn OPS+ into a quartile based metric and still keep the “park” adjustment.  99% would mean that the player (in that park) was better than 99% of players in the league.  I tried to look at this before and just did not see a big enough variation in standard deviation to drastically impact the stats.  The steroid era also had some interesting results if I remember correctly.

On the topic of losing speed, if you shift rightward on the defensive spectrum, I remeber reading a comment Bill James made about Dale Murphy in the 82 Baseball Abstract. I quoted this from the “Baseball Analysts” Abstracts of the Abstracts:

Dale Murphy: “I believe he should stay in the outfield, although possibly in right, and for a simple reason. Talents which are not used tend to deteriorate much faster than those which are tested daily. If Murphy played first base for about two years, he’d wind up a Dave Kingman type, which, physically, is what he is. He would lose the speed and the arm and become unable to do anything other than play first base.”

(Click name)

You could make the argument that moving to an easier position would improve your hitting because the physical demands on defense wouldn’t be as great. Pujols moved from LF to 1B. Of course LF and 1B are right next to each other on the spectrum. In fact Pujols even played a little 3rd base for the Cards his first two years.

Edgar - I don’t know how many of these Tango can get to, but I am taking notes.

I agree with your observation that as a league becomes more elite, the variance from player to player should reduce. I was able to see this in a more extreme situation doing stats for a couple college summer leagues. It’s the top the way hold fairly constant, what changes is how far away is the lower limit of talent.

I do take exception to “steroid era”. Of course some players recently used steroids, but then how many in the past took speed? There is a definite increase in offense in 1993-94. Tango used matched pairs holding batter, pitcher and ballpark constant, and showed that it was an outside factor increasing offense. I did work that showed that even pitchers increased their babip and hr% - the most likely culprit is the ball. There were a handful of “rabbit ball” years - 1973, 1977, 1987 - but this effect has lasted for 15 years. In addition, the whole NL took another spike up in 1999-2001, which did not show up in AL stats. These were Barry Bonds’ peak years, but subtracting him only impactef the league total a slight amount. To some extent he was riding the wave that lifted everyone, and that likely had nothing to do with drugs.

Ball, and/or bat.

Brian, Colin, Pizza, et al: I suggest that you guys contact studes at THT if you plan to do any sabermetric work.  He may include it in the Annual.

I have seen various pieces about some of this, but I have always struggled with random chance versus a chain of events influencing an outcome.  The one that pops right out is the pitch sequence affect - Ball Strike vs Strike Ball.

Simplistic analysis is a pitcher (or batter) has a 50/50 chance of recording an out on the next pitch.  Then you can look at league averages for the situation, pitcher, batter, fielding, umpires, weather, etc. and get expectancy charts, etc.

Perhaps you can look at F/X data to see if strike one is low outside corner, how often is the next pitch an outside slider/curve and what is the result?  Compare this to a low and inside corner strike followed by an outside slider/curve.

How does the catcher influence this?  How does the umpire?  Base-out situation?

Maybe pick a few parameters and see if anything pops out.  “Can we use the data to optimize pitch sequence” I guess would be the hypothesis I propose.

Brian I already posted what I did on lookoutlanding.  I’m not sure I can post a link in here but I’ll try anyway.  I’m not sure its steroids but it seems very convenient.

http://www.lookoutlanding.com/2008/1/11/25150/9196

This is the article Brian is referencing:
http://www.hardballtimes.com/main/article/changes-in-home-run-rates-during-the-retrosheet-years/

It is really the way to do it, since it controls for the exact participants from year-to-year.

I’m not trying to say that looking at variance is the best way to look at the steroids problem.  Just an interesting coincidence when I was looking at variance.

There is so much Pitch f/x to examine too.  A lot of the interesting things is examining the game in detail and what goes on in the head of pitchers and hitters.  Is it better to go outside after a pitch inside?  Many of these articles might not be heavy hitting but are very interesting and actually probably impact how I view the game much more than some of the number heavy research out there.

Something about WPA would be my not so good suggestion.

Post 7 was marked for moderation and is now open.

I am embarrassed to say this because i have a feeling some one has already completed the work and i have missed it:

1. How does moving the infield impact things?

2. Does Coors field impact the number of assists a player has on defense?

3. Who is more successful in making the big leagues?  Minor league relievers or minor league starters?  Is there an advantage to come up as a starter or reliever (allowing for role change)?

4. What impact does the size of the stadium play on DER or UZR or some such defensive metric?

5. Is fouling off pitches a skill or random chance?

6. On a double steal is it more advantagous for the catcher to throw to 3rd or 2nd? 

I have a million more questions, but i am not well versed in what’s out there on the web and in books.  These questions may have already been answered years ago.

I’d like to see an analysis looking at GB-rates for pitchers coupled with infield defensive metrics and pitcher/team performance.  For example, if you have a GB-heavy pitching staff, how does the defensive value difference of individual defenders get multiplied or contribute to team/pitcher performance?

One thing I started working on recently, but was far too lazy to complete, was creating a run expectancy table for a single pitcher, namely Mariano Rivera.

I wanted to see if playing small-ball with Rivera on the mound was the “high-percentage” move, since you can be expected to score less runs against Rivera than against the average pitcher.

I started going through Retrosheet’s play-by-play data, and it just became too daunting a task for me.

Tango, perhaps you could do it not just for a specific pitcher, but for all the “elite” closers as a group?

One topic which understandably came up on our Reds forum is “Should he (aka Dunn) swing?” It’s probably the most common form of fan analysis after balls and strike calls themselves.  But on what are they basing their internal calculus?

Are they factoring in contact rate on given types of pitches?  Are they factoring in batted ball quality and out conversion?  Are the factoring in the value of simply seeing another pitch which could make for an easier swing choice and thus better result?

I had built a little swing calculator in excel based around these and other assumptions.  It showed that with less than 2 strikes, Dunn was significantly better off not swinging unless he was sure the ball was a strike he was quite likely to hit well.  Essentially, if he wasn’t inclined to swing, he shouldn’t.  The argument of the value of putting the ball in play grossly underestimated the impact of swings and misses and poorly hit balls relative to the alternative of getting to see another pitch.

It seemed like the obvious conclusion, but it was very powerful to those who have trouble properly valuing the alternative.

In terms of an article, you could examine the cost benefit of certain types of players swinging at certain types of pitches in certain situations.  Perhaps choose 2 or 3 common situations of interest (full count, start of the PA, bases empty/full) and 2 or 3 different player types (Dunn vs Pujols vs Polanco).  Should they swing at that borderline pitch on 2-1?

Tom/18: I have a generic type of RE table for elite pitchers in The Book. However, it’s a fair point to come up with a “real” RE table for elite pitchers.

Rick/19: that’s been on my todo for several years now.  That’s a project that deserve its own Chapter, a good several dozen hours of research to do it really well, which is outside my current scope.

Keep it coming guys.  Regardless of whether I’ll tackle the issue for the Annual, it adds to my todo, and will give other researchers ideas as well.

I like Bill James line from a long time ago: “I can’t do this alone.” The more saberists out there, the better off we’ll all be.

I always wondered about extreme shifts (i.e. the Williams shift) and if they effect BABIP rates?  I’m a Reds fan and I often wondered if it wouldn’t make sense for Dunn/Griffey to try to go the other way.

Maybe a better question is:  is it bad/good luck when a player has a BABIP at the extreme’s.  We may think it’s random chance, but has there been a study that really proves that?

I would love to do a real study about all of those defensive positioning issues that you guys are bringing up, but unfortunately that data simply isn’t being recorded. You can extrapolate it at the team-season level from hit location data, but I don’t know if that’s enough detail to work with for a study like that.

Which pitchers help them most by working the count in their favor?  I suggest using the WE by count (from http://baseball.bornbybits.com/statistics/WPA.html).  How much do some pitchers rely on getting hitters to good counts?  I think there might be some interesting stuff here and the great thing about using WE would be that it works in leverage/context into the pitcher’s performance.

Tom - #18. Here is an RE table of all pitchers 2005-2007 that had 60 or more saves total for the 3 years.  It is an empirically calculated table on PA’s in the 8th inning or later.  All Away team batting Pa’s are included.  Only Home batting PA’s for the 8th inning are included.

BASE_OUT_STATE PA RUNS RE
0-------------2573-----947---.368
1-------------1915-----378---.197
2-------------1555-----122---.078
30--------------20------22--1.100
31-------------101------80---.792
32-------------187------62---.332
200------------167-----137---.820
201------------359-----194---.540
202------------504-----110---.218
230-------------25------44--1.760
231-------------92------99--1.076
232------------135------46---.341
1000-----------544-----408---.750
1001-----------682-----263---.386
1002-----------664------88---.133
1030------------54-----108--2.000
1031-----------125-----128--1.024
1032-----------181------75---.414
1200-----------132-----198--1.500
1201-----------296-----253---.855
1202-----------436-----132---.303
1230------------42------89--2.119
1231-----------128-----187--1.461
1232-----------154------83---.539

Note the small number of PAs for many of the Base
Out States and adjust for small sample error accordingly. The 1030 BOS is the only one where the value is greater the RE for all pitchers (1.82) though.

Thanks Peter!

Peter that is fantastic!  Even better, because I published in The Book the Markov version using “standard hitting profiles”, we can make a decent enough comparison.

You guys following at home (or via the Search Inside feature at Amazon), go to Table 9, page 31).

In there, I show the RE of bases empty 0 outs as .356.  Peter’s sample is .368, which is fairly close.  We can even play along here:
http://tangotiger.net/markov.html
And change AB to 44, and SO to 9.
Click CALCULATE and you get an RE matrix of bases empty 0 outs of .369.

Lovely, right?

Compare the bases empty, 0 outs.  All three charts are very close to each other.

This is Peter’s man on 1B lines:
1000-----------544-----408---.750
1001-----------682-----263---.386
1002-----------664------88---.133

This is the Markov chain from my site:
0.724 0.412 0.167

And in The Book:
.704 .408 .167

Remembering that The Book values should be a bit less than the other two across the board, we see a fairly big difference with man on 1B and 2 outs.  But with only 664 situations, we have to be aware of sample size.

Peter’s man on 1B and 2B:
1200-----------132-----198--1.500
1201-----------296-----253---.855
1202-----------436-----132---.303

The Markov site:
1.298 0.804 0.373

The Book:
1.271 .766 .350

Peter’s first line is low on sample size, and the third one is as well, but again, with 2 outs and runners on base, Peter’s sample shows fewer runs than expected.

The other one that Peter has with decent sample size is man on 2B:
200------------167-----137---.820
201------------359-----194---.540
202------------504-----110---.218

Markov site:
0.895 0.551 0.269

The Book:
.936 .550 .262

Once again, low with 2 outs with Peter’s sample.  And even low with man on 2B and 0 outs.

At this point, it would be better if Peter can show the overall AB, H, BB, HR, SO numbers.  Maybe the reason they’re not scoring runs with 2 outs with men on base is because they don’t give up alot of HR.

Fascinating…

A couple other thoughts:

Some analysis of one-run strategies.  A one-run expectancy matrix.  An analysis of the correlation between records in one-run games and the run-scoring environment.

PA------AB---H----1B-----2B--3B--HR---BB---K---IBB
11071-9928--2156--1505-392--51-208--786-2789--102

BB are actually non-intententional walks.

If I add 100 reaching base on error, and the 102 IBB back in, the Markov program spits outs .360 RE for bases empty, 0 outs, which compares to Peter’s .368.  If I also knock out 100 AB, I get .368.  Let’s say we go with that, since really we are interested in a “profile” more than anything.

HEre are Peter’s 2 out numbers, with the Markov numbers (using Peter’s input data with my slight alterations) in parens:
BASE_OUT_STATE PA RUNS RE
2-------------1555-----122---.078 (.071)
32-------------187------62---.332 (.301)
202------------504-----110---.218 (.270)
232------------135------46---.341 (.505)
1002-----------664------88---.133 (.167)
1032-----------181------75---.414 (.402)
1202-----------436-----132---.303 (.376)
1232-----------154------83---.539 (.669)

The weighted average (using Peter’s weights) of the 2-out numbers is .188 for Peter’s sample and .215 for this Markov.

Total number of opps is 3816, which is still fairly low.  I’m not sure how significant that difference is.

It’s possible that the closers perform better with 2 outs.

Great stuff Peter…

There were 106 ROE.  I can give you 2 out hitting stats too if you want.

Show the 0-, 1-, and 2- out hitting, with men on base only, since that seems to be what is skewing the results.  Your presentation in post 28 is perfect (if you also include RBOE).

Thanks…

With men on base:

OUT-PA--AB---H--1B---2B--3B--HR---BB---K--IBB-ROE
0--984--818-209-150--30---4--25---50---219---2-14
1-1783-1531-357-253--76--10--18--129--447--44-21
2-2261-2006-382-282--58--14--28--185--586--54-18
T-5028-4355-948-685-164--28--71--364-1252-100-53

With nobody on:

OUT-PA--AB---H---1B---2B-3B--HR---BB---K--IBB-ROE
T-6043-5573-1208-820-228-23-137--422-1537--2-53

HR rate is about 53% higher with nobody on.  Walk rate because of intentional walks is about 31% higher with men on base. Walk rate with men on and 2 out is almost exactly twice the walk rate with men on and 0 outs.

I second #19, although I understand Tango’s objection that to do it right would be a lot more complex than a 4 page article.

One subject which has been discussed is the “what type of team should do best in the postseason?” matter. I recall articles by Kumar in an old HBT, and another one in the BPRO book.

But I think Tango could do a better job of wading thru the murky records and sample size issues, to see if there is really something there, both in an actual sense and in a theoretical sense.

I think I’ve seen recent posts by Tango and MGL that there’s not much to this.

IOW, I guess, everything is pretty much dwarfed by the random element in a short series, and furthermore if a team focuses on some limited types of ballplayers in their acquisitions, they might have a great postseason-style team, but might not in fact make the playoffs.

Still, I find this an interesting topic, and would love to see a clear examination of what should work in theory.

I forgot to mention the B James W Series prediction system, published 25 years ago. So, this is a topic which has been around for a long time. And each researcher looks at the unusual results, and gives his interpretation of why it works like that.

I wanna see this subject Solved (as in Solving DIPS).

I’m not nearly as hard-core as most of you guys, so forgive me if the following are unworthy, but here are questions I would find interesting:

1) How does a player’s offensive profile change in value depending on the caliber of the opposing starting pitcher?  This might have an impact on the question of postseason results, where teams will generally face top-of-the-line starting pitching.  i.e. given that the overall run environment will be lower, do HR hitters become more valuable, because of the presumed difficulty of stringing hits together?  Or if the pitcher is good at suppressing HR (I know that many ace pitchers are actually FB pitchers, but hypothetically) does this make the HR hitter less valuable?  This should be pretty easy to do, if it hasn’t already been done, but fully exploring the change in lwt values of different events based on the individual pitcher’s tendency to inflate/suppress the events, as well as the overall run environment, might be an interesting exercise in any case.

2) How about a study on the hit-and-run?  Unless I’ve completely overlooked some studies out there, this seems to be a sorely neglected topic.  How much likelier is a GB to become a hit when a runner takes off from first?  With this in mind, what values of the different variables involved (expected contact percentage/GB rate from the batter/pitcher matchup, speed of the runner on first) generate a break-even point for the hit-and-run in different base/out/count states?  Do major league managers do a good job of using the h/r in the proper situations?

Last year was the first time I bought the annual.... can you give us a list of previous articles you’ve done for the annual with a very brief description? Your WOWY articles were a different style than your writing here, so I’m having difficulty imagining what to suggest.

IOW, I don’t think you want to do an “exploratory” article, such as the articles written daily on the THT website.

What about looking at maximizing RE each inning vs. maximizing the chance of scoring at least one run each inning?  Perhaps in different run environments?

Peter in post 32, converted to per 600 PA (excluding IBB):

OUT PA AB H 1B 2B 3B HR BB K IBB ROE
0 601 500 128 92 18 2 15 31 134 1 9
1 615 528 123 87 26 3 6 45 154 15 7
2 615 545 104 77 16 4 8 50 159 15 5

As you can see, an enormous drop in HR with 1 and 2 outs, and so, that by itself tells me why Peter’s data is showing fewer runs than expected when 2 outs.

Typically, 10% or 11% of your hits will be HR.  Probably a bit lower for great pitchers.  The numbers for these closers is 12%, 5%, and 7% with 0,1,2 outs respectively.

The 1-out number is very statistically significant.

There is a good potential for research on how batters/pitchers change their approach based on the game state.