THE BOOK--Playing The Percentages In Baseball

Monday, July 02, 2007

Sports Economic Papers

This was sent to me recently.  I haven’t looked at any yet:
http://ideas.repec.org/s/spe/wpaper.html

Friday, June 29, 2007

Bad research by someone who probably should not be doing research

On ESPN.com, there is an article by Jayson Stark which, among other things, talks about a phenomenon known as “World Series Burnout.” I am not sure I have ever heard of it, but basically it postulates that pitchers who pitch in post-season play (I am not sure why only WS pitchers are inlcuded other than the fact that they had the opporunity to pitch more in the post-season) experience a dropoff in performance and innings pitched the next year, presumably due to extra stress on their arms (I guess).  Now, I am not saying that this is not true (it could be I suppose), only that the evidence and conclusions presented by Stark are poor.

Here is part of what Stark says:

Those White Sox were about as powerful an argument for the existence of World Series burnout as we’ve found. But we went further. We studied this phenomenon ourselves, with the help of the ever-inquisitive folks from Baseball Prospectus. Here’s what we discovered:

• Just eight times out of the 47, on the other hand, were pitchers able to lower their ERAs by half a run or more. And the pitchers for all of those eight occasions—Greg Maddux ( twice), Andy Pettitte (twice), Tom Glavine (2000), Roger Clemens (2004), Livan Hernandez (2003) and Matt Morris (2005)—were pitchers who had been through October before.

FIrst of all, the second and third bullets are essentially the same as bullet #1. No worries yet though.

Baseball Prospectus’ Bill Burke, meanwhile, studied all World Series staffs from 1996 through last year’s Tigers and Cardinals. Then he charted what happened to every pitcher the next year, even if that pitcher changed teams. He found those pitchers got worse in every major rate category the following year. Take a look. (Each category is “per 9 innings pitched.")

Now granted, those aren’t huge drops. And we are talking about comparing seasons where, in general, everything goes right with seasons where life returns to “normal.” So maybe this is just a predictable drop-off in many respects.

Actually, Stark hits the nail right on the head.  Since WS teams on the average have above average (probably well-above average) pitching and hitting, they will all experience normal regression to the mean the next season, in both hitting and pitching.  In fact, I suspect that you would see almost the exact same results if you looked at hitter stats (although hitters don’t regress as much as pitchers given the same number of PA or BF).

Now, I have not calculated whether the above regression is more than you would expect from “normal” regression toward the mean.  It certainly looks “normal” to me.  If someone wants to do that, remember to NOT use the total TBF or IP for the staff when figuring how much to regress.  A team of 10 pitchers where each pitcher throws 100 IP is NOT the same as one pitcher who throws 1000 IP, from the standpoint of regression toward the mean.

Anyway, back to Stark:

But when you look at last year’s White Sox, when you look at what happened to Chris Carpenter in St. Louis this year, when you look at a Tigers team that went from having the most unhittable bullpen in the American League (.242 opponent average) to the third-most hittable (.275), how can anyone say World Series burnout is a myth?

So he then dismisses the probably correct conclusion (normal regression toward the mean) because of one team team and one pitcher (anecdotes)! Arghhh!!!

At the very least, one would want to compare the above data with that of a control group, say all non-post season teams that had around the same pitching stats.  My guess is that you would not find any differences between the control and experimental groups.

Or, as I already mentioned, you can adjust the above data to account for normal regression toward the mean and then see if there is anything “left” (more performance decline).  Since it can be a little tricky properly accounting for normal regression, the control group idea is probably the best way to go about it.

Wednesday, June 27, 2007

All things Del Grande

Billy Beane responding to the Del Grande article:

http://www.insidebayarea.com/sports/ci_6240002

Blogging of the Del Grande article:
http://sports.aol.com/fanhouse/2007/06/25/did-race-play-a-role-in-the-milton-bradley-situation/

Mailbag of said newspaper on the Del Grande article:
http://www.insidebayarea.com/turn2/ci_6240003

Comments of said newspaper on the Del Grande article:
http://www.haloscan.com/comments/insidebayarea/6223509/

A fellow reporter of said newspaper on the Del Grande article:
http://www.insidebayarea.com/sports/ci_6239999

The current list of Del Grande articles:
http://www.insidebayarea.com/columnists/davedelgrande
No single mention of any article being removed.

The Del Grande article:
http://www.insidebayarea.com/athletics/ci_6223509
Uhhhh.... it’s been taken down.  Ok, never mind.  It never happened.

However, here’s the Google cache of the Del Grande article:
http://72.14.209.104/search?q=cache:pJC7RIdJQV4J:www.insidebayarea.com/athletics/ci_6223509+%22Dave+Del+Grande%22+%22billy+beane%22&hl=en&ct=clnk&cd=2&gl=us

Grab it quick, before it’s removed.  And because it is no longer part of our history, we are doomed to repeat his tirade by others of his ilk.

Sprint Ambassador

I was a Sprint Ambassador for six months.  For those who don’t want to click the link, Sprint sent me a free phone with free premium services (including NFL mobile, that I never used) because I was recognized as a “top blogger”.  I don’t know how they came to that determination, but I’m guessing there must be someone in the Sprint marketing department who must be really into sabermetrics. My wife and I were/are coincidentally actually already Sprint users for several years, so we’re predisposed to give them a favorable rating.

Anyway, the program itself, while a good initial idea, was poor in its execution.  There really was no two-way communication.  I wrote some comments once, and I got the feeling it went into the internet abyss.  Plus, the phone had so many features, I didn’t really want to learn too much of a phone that I had to give up in 6 months.  And even if let’s say I wanted to use all the services, I had no idea how much anything cost, since the bill was redirected to their marketing department.  So, say I really like the services, and decided that, ok, I’ll join.  Wouldn’t it be nice to know that I spent say 99$ a month?  As it turns out, all we ended up doing was using our 3rd phone for pictures, MP3, and calling Canada while on the road.  We weren’t going to give anyone else a phone number for 6 months, and we already had Sprint, so all we ended up doing is splitting the minutes over 3 phones instead of 2.  I’d bet that out of all the bloggers, our phone use was dead last.

The presumption of Sprint is that their design team and focus groups created a fantastic product, and it was up to us, the bloggers, to spread the word.  But, the phone and sevice, while good of what I used, was not the fantastic product that was Sprint’s presumption.  In short, they didn’t use bloggers on the right product/service.  They could have used us bloggers as a focus group, and set up a forum for us to blast away.  That didn’t happen.  What should Sprint do?  Recontact all us bloggers, invite us to a forum, and ask us the best way to use bloggers.  Sprint’s method of throwing things out and seeing what sticks is an act of cluelessness.  A study requires controls and identifiable parameters.

Tuesday, June 26, 2007

Intentional Walks as Leadoff Hitter

One of the remarkable things about Tim Raines is how much he was intentionally walked.  As a leadoff hitter, he was IBB a career total of 102 times, with a career high of 14 in 1987 (the year he missed the first month due to collusion, and actually batted 3rd often enough that he only started 58 games in leadoff!).  From 1985 through 1988, he was IBB 43 times in the leadoff spot in 336 starts.

Here are 93 of those games (including multiple IBB games), thanks to Play Index.  He actually had a stretch of 5 straight games with an IBB as the leadoff hitter of the game.  I can’t tell if Raines is the career leader in the category of lifetime career IBB in the leadoff slot, but I’d sure bet on it.

Inferring Injuries

When you establish the true talent level of a player (or create a forecast, which is essentially the same thing), you want to know if he is healthy or not.  And, if he’s not healthy, how unhealthy, and how persistent is his illness.  So, you try to infer such things.  If a player plays 159, 160, 154, 159, 161, 112 games, his OPS+ at any point in that stretch bottomed at 133, and the player is 27 years old at that point, we infer an injury. We don’t need to look more closely at the situation, though it could have been the case of someone even better usurping his playing time.  You always have a certain uncertainty level.  And if the player is 37 instead of 27, we may be more inclined to infer a longer rehab period.  But, we still don’t know what kind of injury because the data doesn’t tell us much more.

John Walsh shows us the data for Curt Schilling.  Now, we don’t need to infer if his performance was about balls falling in for a hit, or whether his true talent level was marketdly different.  We remove that uncertainty level with the data.  Depending on the nature of the illness, we’ll be able to either discount the data from this performance more, or place a greater premium on it.  We’re always looking for the establishment of a new talent level, as opposed to randomness creating noise around the data.  It’s data like this that we need.

And for MLB teams that are not doing this.... are you kidding me?  What Walsh, Fox, Beamer, Sheehan, Appleman, et al are doing is the cutting edge of sabermetrics, the point where performance and scouting converge.  This is the pot of gold that is being prospected.

***

Further research would go into the “mix” of pitches, and the “strategy” of pitches, based on the game state (inning, score, base, out) conditions… i.e., Leverage Index.

***

The data itself also has a certain amount of uncertainty, as can be easily seen with David Wells having a bunch of pitches being released from the wrong side of the mound (four feet from where it should be). 

Thursday, June 21, 2007

Translating homeruns

Patriot gives us a good introduction on translating (rescaling) stats, focusing on the HR, which if you cut right to the chase is based on this:

New HR = HR/(PF*RPG)*9

In essence, Patriot is scaling it linearly to runs per game.  So, 25 HR in a 2.5 RPG environment would scale to 50 HR in a 5.0 RPG environment. If we run the Markov calculator:
http://www.tangotiger.net/markov.html , we see that 0.66 HR hit in a 2.5 RPG environment (set AB=51, or multiply all the default numbers by 27/41) would be equivalent to 1 HR in 5.0 RPG.  (Note: the run value of a HR, while fairly stable, drops by about 5%.)

What if we go to win values instead?  Using PythagenPat, the win value of a HR is .133 wins in a 5.0 RPG and .21 wins in a 2.5 RPG environment.  That 1 HR in a 5 RPG environment is worth .133 wins.  And, how many HR in a 2.5 RPG would be worth .133 wins?  0.63 HR.

As you can see, both approaches give a fairly similar number, and is a bit different from the 0.50 HR that Patriot would propose.  However, I chose rather extreme environments, and perhaps in more realistic extreme environments, we won’t find such differences.  Trying a 3.5 RPG environment, Markov gives the equivalency as 0.82 HR, and PythagenPat says 0.79 HR.  Patriot’s approach would have said that 0.70 HR in a 3.5 RPG environment would translate to 1.00 HR in a 5.0 RPG environment.

Walks: Of learners and idiots

Similar to what I did with strikeouts, I now look at walks.

UPDATE (Jun 21, 14:35): I misreported the direction of BB/K in the article.  It has now been updated. 

BIS at THT

An excellent primer by studes on the latest fielding stats available at Hardball Times.

Wednesday, June 20, 2007

Draft Picks or Young Players?

Dave at USSM looks at what to do with Ichiro: trade him for current players, or wait to claim draft picks?  He rolls up his sleeves, does all the great dirty work that we hope for, and finds this:

Out of the 41 players received in return for the 16 traded all-stars, two have turned into all-stars (Grady Sizemore and Aaron Harang), several more are good everyday players or mid-rotation starters (Mark Teahen, John Buck, Brandon Phillips, Placido Polanco, Cliff Lee, Jake Westbrook), and the other 33 aren’t in baseball anymore or have little to no value…
Of the 42 compensation selections on the list, three have become all-stars - David Wright, Nick Swisher, and Huston Street. Several more have become solid contributors - Joe Blanton, Mark Teahen, David Aardsma, and Aaron Heilman. And, a few others have become the elite prospects in baseball today - Philip Hughes, Jarrod Saltalamacchia, and Adam Miller, while JoJo Reyes is just a good pitching prospect instead of an elite one, and ‘07 draft picks Beaven, Borbon, Smoker, and Zimmerman are far too young to determine their value at this point.  27 of the picks could be labeled as busts, even though a couple still have a shot to turn into major league role players down the road.

I don’t like the “aren’t in baseball anymore” provision, since they could have certainly done something from then through 2006 (Bobby Kielty for one).  However, Dave does give you the list of names, so you are free to make up your own conditions.

At the very least, the “get something for him now” crowd clearly aren’t thinking about the opportunity cost.  There’s this thinking that if you can’t see it, it doesn’t exist.  And draft picks are treated as some distant possible hope, a lottery, by this group.  On the flip side, even if the draft picks return better players than the “get something for him now” players, you still have to wait a bit to get those players.  You need to apply some discount value to that.

In the end, dropping Ichiro means that you lose him for two months and you lose the two draft picks (one a bust, and one who may be ok), in return for three players, two of which will be a bust, and one of which may turn out ok. 

Pick your poison.

Tuesday, June 19, 2007

Basic Aging Curve for Hitters, 1957-2006

I updated my aging curves, this time for the time period 1957-2006.  Here’s the step-by-step process:

1. Select only nonpitchers
2. Total on: player, year, team (i.e., Abreu gets two records in 2006) ...27,497 records
3. Determine the main home park for each team, and append that to the record (Abreu gets Citizen’s park for one record, and Yankee Stadium for the other record)
4. For every consecutive years, match on player, and main home park.  (This is important for guys who play on the same team, but switch home parks, like Pujols) ...15,715 paired matches
5. Figure the age for each player as season minus birth year
6. For each match, give a PA weight based on this formula:
2 / (1/PA1 + 1/PA2)
7. For each age, sum up the PA weights of all players, along with their wOBA in the two matched years.

You get this Google Docs

Legend:
Age1: the age of the first of the two matched seasons
n: number of players in the sample
PA: the total of the weighted PA
wOBA_1: the weighted wOBA at Age1
wOBA_2: the weighted wOBA at Age1+1
diff: the difference between the two wOBA; a positive numbers means performance improved
ratio: the odds ratio, which can be loosely described as wOBA_2/wOBA_1; a number above 1.00 means performance improved

I then added another column called wOBA_curve.  This value is based on the ratio (or diff), and allows us to create a single curve.  For example, at age 22, we see .332, and at age 23, we see .343.  That’s an 11 point improvement between 22 and 23.  If you look at the “diff” column for age 22, you’ll see a 10 point difference (consider it rounding error).  The same player, at age 23 (.343) becomes .345 at age 24, or a 2 point improvement.  The “diff” column for age 23 shows a 3 point improvement.

This is called Chaining, and it allows you to take the known differences of samples, and merge them onto a single player.

You will note what I did not do (yet): regression.  You will notice that the values in wOBA_1 constantly increase from age 22 through age 40.  However, starting from age 27, the PA goes down every year.  This is survivor bias: the only guys left in the sample are those guys deemed good enough to still be there.  And, teams obviously have an age bias.  That is, of the guys at age 40 who are allowed to play at age 41 (only 39 of them) are BETTER than the guys at age 25 and allowed to play at age 26 (1652 of them).  Of course, if we only looked at the 39 best 25 year olds, they’d be far better than the 39 best 40-yr olds.

Anyway, back to regression.  As you get older, teams will more likely depend on how you performed recently than on any “true talent” skill you may still have.  So, there is an enormous selective sampling issue.  As you can see by the year 2 performances, every single age class from age 30 onwards performed at around the .333 wOBA level.  The old man dropoffs you see in the late 30s and early 40s are biased.  The year 2 performances are more indicative of the talent than than the year 1 performances.

If you don’t apply some regression, what happens?  According to this basic chart, a 19-yr old is equivalent to a 34-yr old.  Seeing that there are 10 times more 34-yr olds and 40 times more PA given to 34yr olds, compared to a 19-yr old, that seems like a ludicrous statment to make.

I also didn’t adjust for the change in year-to-year environment, like the juiced ball or mound raised.  However, we don’t expect to have a disproportionate number of players at a certain age group in 1968 or 1987.  So, this chart would be unaffected. 

If there is a systematic bias, like the talent pool keeps increasing each year, and therefore, the difference in the paired ages captures not only the change in performance, but also the increase in talent, then this aging curve will make it seem like players age slightly faster than they actually do.

At some point, I’ll repeat this for each of the “skills” (like BABIP, K/BB ratio, HR/XBH, etc), and I’ll apply regression.  Until then…

Do you know how much Selig makes?

14.5 million$!  That puts his salary tied for 15th in 2006 of all MLB players!  Bettman is behind 22 NHL players.

I remember Scotty Bowman, one of the greatest sports coaches in North America ever, said that his salary should be set around the same as a 3rd line player.  (In the NHL, there are 4 lines, so the 3rd line would be a bit below average, and the 4th line being a nick above replacement-level.) I think Marvin Miller set his salary as the average MLB player.  I think Fehr continued that tradition.

What’s amazing about Selig, is that not only is he #15 of all players, but unlike players, he doesn’t have a limited shelf life.  He could, theoretically in his life span, earn more than any MLB ever, except for a handful of players (ARod, obviously).  Through 2006, Bonds has earned 173 million$ in salary, followed by ARod at 148 million$.  I don’t know how much Selig has earned since he was commish (I’d estimate close to 100 million$, assuming a constant 15% raise, culminating to his current 14.5 million$).  Assuming he’s got two more years, he’ll end up earning around 130 million$.  If Roger Clemens didn’t come back this year, he’d have earned more than The Rocket.  Selig is definitely in select company.

Unbelievable.

The Replacement Pitchers

Sean chimes in:

How good (or bad) is a replacement level starting pitcher?  I tried to answer this by looking at all starting pitchers for 2007, and ... take out all pitchers who started the season in the rotation, those who would have started had they not been injured, top prospects (those who made Baseball America’s top 100 list), and Roger Clemens.

THAT is about as perfect a definition of replacement-level starter as there is.  Note, this is NOT the same as a replacement-level pitcher in a starterRole.  After all, where do you find a replacement-level starters?  In the bullpen!  And those guys are (slightly) better than the scrap heap.

How are they doing for 2007?  After taking out the non-replacements, as well as any pitcher who made more than half of his appearances in relief, I have 952 innings with a 5.03 ERA, against a league average of 4.32.  That’s a winning percentage of .430

I don’t get the “relief” condition.  Personally, we should stick with only performance in starterRoles.  Fangraphs and Baseball-Reference both give you the necessary breakdown.  In any case, as I’ve shown in the past, here’s our benchmarks:
- replacement-level pitcher in a starterRole: .380
- replacement-level pitcher in a relieverRole: .470
- average starter in a starterRole: .490
- average reliever in a relieverRole: .520

As you can see, the relief-to-start adjustment is about -.090 wins.  So, an average reliever in a relieverRole (.520) would translate to .090 less than that in a starterRole, or.... .430. 

Ace relievers are not candidates for a starterRole, and if we assume that they make up 1/6th of the relief pool, and, in a relieverRole, they are .600 pitchers, then the .520 breaks down as: .600 for ace relievers, .500 for average reliever, all in a relieverRole.  That .500, subtracted by .090, gives us .410 for non-ace relievers in a starterRole.  Therefore, my best guess as to what a replacement-level starter looks like (guys in the bullpen who are not ace relievers) is a .410 pitcher in a starterRole.

Because Sean’s pool of pitchers is overweighted by Guthrie, I’d bet he’d get numbers around .410.

Friday, June 15, 2007

Batted Ball Breakdown

The always resourceful Dan Fox gives us a breakdown of batted balls.  Here’s the summary, with a little change on my part:

HITS. SLG.. LWTS w/GIDP Type
0.728 0.940 +.324 +.324 Line
0.177 0.302 –.126 –.126 Fly
0.241 0.262 –.108 –.126 Ground
0.020 0.024 –.284 –.284 Pop

Those numbers EXCLUDE homeruns in the numerator and denominator.  So, a flyball is a hit only 17.7% of the time, but the slugging average on a flyball is .302.  Groundballs have a hit .241 times, but with a slugging average of .262.  As it turns out, the extra oomph on extra bases for flyballs can’t overcome the extra outs for a hitter.  The Linear Weights run value of a flyball (excluding HR) is about .02 runs lower than your standard groundball.  However, don’t forget the GIDP.  Throw the GIDP in, and guess what?  The run value of a FB and GB are virtually equal!

But remember, the reason that we prefer GB pitchers, generally speaking, is that they give up less HR.

The run value of a line drive is similar to that of a walk, and the run value of a popup is similar to that of a strikeout.

Thursday, June 14, 2007

The Complete Dewan Fielding Plus/Minus

It seems that Dewan will see MGL’s UZR of end of May, and raise him a Plus/Minus of mid-June.

Note to ESPN: this is what you should be doing. 
Baseball Prospectus: you too.