THE BOOK--Playing The Percentages In Baseball

Well, what I said was that I “constantly disagreed” with you. Which was true: at the time we were debating whether or not MLB could sustain current revenue growth indefinitely, whether fan scouting reflected media bias, finer points of community projections, and something else. It’s not as if I thought your arguments are meritless or anything.

That’s the way I took it, like a healthy debate where two people spar, each side not really swayed by the other, even though the positions have merit.  I’m at a disadvantage though, since you can weave words better than most.

JC’s book is good.  I can’t remember exactly why as I would have to skim through it again, but it is definitely worth reading.  The other similar one (is it by Gennaro?) is very good also.  In fact quite good - must reading for a GM (at least a GM who is concerned with the team’s bottom line - and not just winning at any cost or pleasing the fans, players or media) or owner.

For those interested, here is an interview I did with Derek about his new book:

http://www.hardballtimes.com/main/article/coffee-with-derek-zumsteg-the-cheaters-guide-to-baseball/

Quite a backlash against Derek on this one over at Primer.  But, Derek needs to be judged in context.  For example, if we check out Alexa.com, we see that my blog (this one right here), reaches .0005 percent of the internet users today.  That’s 1 out of every 200,00 internet users.

Derek’s blog reaches far less, though on Apr 4 he jumped to one out of every 50,000 internet users.  USSM, his other blog which he shares with Dave, reaches around one out of 100,000 internet users.

Baseball Prospectus reaches one out of every 25,000 internet users.

Now, how much responsibility do I as a blogger have?  Can’t I say here that it looks like Kenny Rogers doctored the ball, and should be suspended?  Can’t I say that I believe that Clemens intentionally hit Piazza in the head?  If I start to present some evidence of any of that, and make a biased observation to support my theories, is that wrong?  What I’ve got here is a rinky-dinky little blog.  USSM’s blog is fairly well-supported, and I’d be ok if they did it too.  If it wasn’t for blogs, we’d have never been able to accuse, try, convict, and later plead mea culpa against Antonella Barba.

Baseball Prospectus has a certain reach and they don’t treat their product as a blog (weblog, or web-log, which is a journal of your thoughts posted online).  If they presented what Derek did in the manner in which he did, then maybe you have a case.

This is why we have blogs, as a starting point to exchange ideas, where further research is warranted.

If you have a problem, take it up with the national media that picks up the story and treats a blog as something more than it is.

That said, any publicity is good publicity, and regardless of whether Derek is right or wrong, he must be somewhat happy with how things are playing out.

Ok, finally got both books.  Any suggestions as to the first chapter in JC’s books that’s worth getting me into a sabermetric frenzy (beit good or bad).

Next up on the hit list:

http://www.zumsteg.net/cheatersguide/2007/04/18/joe-nathans-pine-tar/

I’m going to presume Nathan is doing nothing wrong given Zumsteg’s shameful rumor mongering on K-Rod, which has been quite soundly debunked.

Zumsteg strikes again:
http://www.zumsteg.net/cheatersguide/2007/04/21/a-bubbly-reminder-of-a-corrupt-past/

And the Twins respond:
http://www.zumsteg.net/cheatersguide/2007/04/24/royals-return-champagne-mlb-pokes-around/

Derek will soon become the most hated person in MLB clubhouses.  Luckily any publicity is good publicity.

I’m reading Derek’s book on and off.  So far, the most enjoyable part of the book is the “Fan participation”.  It’s a good fun read, with a few “sidebars” for good measure.

The parts of the book that fall a bit flat are the “recent history”.  I’m a big hidden ball trick guy, and I’ve seen tape of Dave Bergman do it twice, and Ozzie Guiilen being nabbed twice, and there was no mention of these guys.  And of course, when you put these two guys together, Bergman nabbed Ozzie on one of these plays.  The last paragraph here has the details:
http://www.hardballtimes.com/main/article/baseball-smarts-the-best-percentage-players-in-the-game/

So, I find the book most enjoyable in the pre-80s recaps, especially in the really olden days.

Just read the steroids chapter.  A well-written, and both concise and extended, view of steroids.  One fascinating topic was handled sabermetrically.  In The Game of Shadows, they talk about Bonds being on a 3-week on, 1 week-off cycle, and that his observations shows that he wouldn’t feel well when he’s off.

So, what does Derek do?  He looks at Bonds’ performance in 2002, in clumps of 4 weeks, and… he’s able to find when Bonds’ cycle starts, and he supports Bonds’ alleged assertions that he was hitting not as well.  And the results are rather dramatic.  He also got Woolner to check Bonds in 98, and there was no similar pattern.

This is an enormous finding.  I’ve got my PBP database well set up for 1999-2002, so I’ll run it at some point this weekend to see how this looks, year-by-year (unless one of you guys wants to do it).

This is one of the problems with books.  It gets very little traction, since it’s not linkable.  I hope that Derek would provide an online excerpt to these 3 pages, as I think it’s incredibly fascinating.

I just replicated Derek’s 2002 study of Bonds, and I confirm the enormous difference.

During his three-week “on” cycle, his wOBA is .578, while on his off cycle, it was .445.  Now, I have to first say that his off cycle, as much as it’s far lower from his on cycle, is still a fantastic total!  Pretty much league-leader probably.

Anyway, the off cycle had only 127 PA.  1 SD = 50 points.  So, the difference, 133 points, while enormous, is a bit under 3 SD. 

His HR rate (HR per PA) is 9.8% on his on cycle, and 3.9% on his off cycle.  Also 3 SD difference.

I’ll come back later tonight to see if there’s a pattern in the 1999-2001 years.

In 2001, days 24 through day 30 was his off cycle.  His wOBA on his apparent on cycle was .569, and his apparent off was .427, for a .142 point difference.  This based on 132 PA.

HR difference not as large, with 12.2% of his PA as HR in the on, and 9.1% (!!!) in his off.  Like I said, simply a remarkable hitter, even when he’s not allegedly feeling well.

In 2000, days 1 through 7 is the alleged off cycle.  wOBA of .507 in the alleged on, and .359 in the alleged off.  That difference is 148 points, on 179 PA.  That’s over 3 SD difference.

HR rates of 9.4% and 6.1%.

In 1999, days 13 through 19.  .452 alleged on, .332 alleged off.  Difference of 120 points, on only 90 PA.

HR rate of 9.9% against… wait for it… 1.1%.

Add it all up, and you’ve got some pretty damning evidence here.

***

Thanks Derek.

.138 wOBA difference on 528 PA.  That’s around a 5 SD difference.  Gulp.

Looks like I’ve got to expand my PBP database backwards, and see if I can find other cycles like this for Bonds, pre 1999.  Since we’re pretty much assured that he was clean in those years, if I can find these kinds of cycles, maybe it’s just pure luck.

Also have to look at Jason Giambi, Jeremy, Matt Lawton, Palmeiro, and whoever else to see if anyone else has this 3-week/1-week cycle.

Simply fascinating.

This is really interesting stuff.  One question:  were the start dates for Bonds’ “bad week” generated independently, based on info obtained by the reporters?  Obviously, if there’s a possibility those dates were selected with Bonds’ bad performance in mind, that changes things.  (Although, if the gap remains even when you exclude the first bad week, the results would remain very impressive.)

The start dates were not known independently.  When I read this chapter I thought “selective sampling at its finest”.

What they did was run every possible combination of 3 weeks on/ 1 week off until they found the one with the biggest difference, so of course you are going to find something.

What I don’t know is if the gaps for Bonds are so great that they are meaningful after accounting for the selective sample.

Ah, that’s another matter altogether.  I don’t have the statistical chops to answer the significance question.  But I’d be much more impressed with these results if, say, they held up when you took the worst 7-day stretch in the first 28 days of each season, and then ran the sequence from there.  (On the other hand, I would tend to trust Woolner to be careful when doing something like this.)

No, they did what I did.  Pure selective sampling, and look for the best 3-week, worst 1-week. 

So, in 1999, that’d be days
13-19
41-47
69-75
...

So, the question to ask is: how likely is that you can find such gaps in performance (3 SD of the sample difference).

That’s why I want to run it for Bonds pre-99, and for all players.

That is, assume that every single player is on PED, and selectively sample the dates to fit that criteria.  After doing all that, does Bonds’ performance still stick out like a sore thumb.

The way Derek reported Woolner’s findings for the pre-99 data was that he wasn’t able to find such a big gap.

Another way to put it is this: can I find some hitter’s Thursday performance totally out-of-whack of his Fri-Wed performance?  Clearly, I will find someone with some gap.  How big is this gap that I will find?  Since we have no prior expectation that a hitter will perform better on Thursday than any other day, this means that, regardless of the result, it was the result of luck.

It would be interesting to see if Bonds has a gap like that for 2006, the first year he played under strict steroid testing.

Assuming this isn’t something you could find just by dumb luck for every hitter.

As far as the selective sampling goes, Is this any different than pciking out a hitter’s worst month?

You’d see a bigger split with worst month, since its a hitters worst 1/6 of a season instead of 1/4, but if we assume its random we should see similar results.

Right, you can call Bonds’ Aug 2002 as a “juiced” month:
http://www.baseball-reference.com/pi/bsplit.cgi?n1=bondsba01&year=2002

Apr 2001 might be his “unjuiced month”:
http://www.baseball-reference.com/pi/bsplit.cgi?n1=bondsba01&year=2001

So, the splits that Derek has observed, based on cycles, may simply be the artifact of selective sampling.  That’s why we need to see how all the players in the league look.

Rally/22:  It’s not quite the same as worst month.  It’s not likely that a given player would demonstrate the same 3-on/1-off pattern over several consecutive seasons, IF the pattern is at least somewhat unusual in the first place (which Tango will figure out).  So it’s more like having the SAME bad month four years in a row.  And if a guy was -.140 wOBA in 4 consecutive Aprils, we’d think he had trouble starting the year.

Actually, it is.  Bonds is not showing the same 3 on/1 off pattern every year.

In 2001, off is day 24-30

2000: day 1-7
1999: day 13-19

Months are easiliy available from B-ref, but a better test would be to divide the season up into quarters.  Look at the difference between your worst quarter and the average of your other 3.

And we should look at relative wOBA instead of absolute difference, since Bonds is so far ahead of the league his absolute differences are going to be greater than other players anyway.

Guy/24: I thought that was the case, but thinking more about it, I don’t think so.  I made no attempt to link the 28-day cycle, year-to-year. 

The 1999 off-cycle started 12 days after the 1st game of the season (game 1 = day 1, therefore cycle started on day 13).

In 2000, it was day 1.

In 2001, it was day 24 (or, equivalently, day minus 4).

In 2002, I forget what it was, but it was something like day 20 or 21 (or day minus 7 or 8).

A calendar year has 365 or 366 days, meaning a 28-day cycle gives you 13 cycles (364 days), plus a day or two.  If Bonds was allegedly on PED, and allegedly on a cycle, and he, for whatever reason, maintains the cycle in the off-season (sounds like birth control pills), then if the off-cycle started on Apr 3, 2000, then we’d expect an offcycle in 2001 to be on Apr 2 to Apr 8, 2001.  The data suggests Mar 28 - Apr 3.

In 2002, we’d expect Apr 1-7, 2002 for the off-cycle.  The data suggests Mar 26 - Apr 2 or so.

I mean, I suppose you can argue that Bonds had a lapse (vacation) in his cycle, of the exact same duration between 2000/2001 and 2001/2002.  But, then, explain 1999/2000.

I think it’s dangerous to try to link them year-to-year, unless you have a prior expectation.

Therefore, the first line in post #16 has to be considered as useless.  I have no reason to just add them all up, and apply an SD, I don’t think.

I’m convinced—no need to link the years. 

BTW, David G. said this about Zumsteg’s analysis over at THT, which sounds like less selective-sampling than we’re saying here:  “Looking at game-by-game data, Zumsteg finds those cycles with the help of Indians statistical analyst Keith Woolner. Starting with the first day of each season in which Bonds used steroids according to Game of Shadows, Bonds showed clear patterns of three weeks of great play followed by a huge fall-off in power followed by another three weeks of hot hitting, and so forth. In seasons prior to that, Bonds showed no patterns like this.  Basically, Zumsteg finds that if Bonds had not been using steroids, he would have hit half as many home runs, giving him a still-impressive career total, but leaving him well behind Hank Aaron.”

Even if Bonds’ pattern is highly unlikely to be random, there’s still the question of interpreting the results.  Rather than concluding that Bonds doubled his HR output in these years, the data could also show that Bonds managed to suppress his production one week in four by making himself sick.  Maybe he would have hit 83 HRs without the steriods!

I don’t see that quote refuting the selective sampling.  His alleged PED years are 1999-onwards, and Derek’s table in the book shows that he selectively sampled the off-cycle week.  I can’t attest to what Keith did in the pre-99 time period.  Perhaps Keith didn’t do what Derek did?

I’m fascinated regardless, and so, I hope to continue the study by next weekend.

(The results of the Community Forecasts have to wait yet another week.)

I looked at 2002, and I took various off-cycle weeks.  Bonds for example had the off-cycle week of day 23-29.  For that particular week, he wasn’t even the worst in his group.  Richard Hidalgo and Bobby Abreu had a bigger gap.  Then I looked at other off-cycle weeks (day 1-7, day 8-14, day 15-21), and there was always one or two guys in the nearly 300 players who had a 130 point difference or higher.  It’s fair to say that if I ran it for each of the 28 possible cycles, that I’d find about 40 players who did something similar to Bonds or worse. 

I think what Derek may have found was purely a function of selective sampling.  I’ll study it a bit more.

Tango - I think you are probably right about the selective sampling, but it is possible that you identified 40 other players that used steroids on a 3 week on/ 1 week off cycle.  Maybe you should go back to the early 1980s and see if similar cycles existed in the pre-steroid era.

Peter: right, I should look at the spread of differences, and see if it’s higher.  However, are all PED users on a strict off/on cycle?

One thing that struck me is that if 15% of players are expected, by luck, to have a greater than 130 point difference in any given year, then the chance that the same player doing that for 3 straight years would be .15^3= 1 in 300.  Bonds could simply be the one outlier.

Of course, I selected the 130 point endpoint after seeing the Bonds results.  If I drop it down to 120 points, and expand to 4 years, then it’s .50^4=1 in 16, and it becomes much less impressive.

I’m thinking there’s really nothing here, other than a happy coincidence.  I’m going to run a full-report on all players from 1999-2002, picking out the off cycle for each one, and we’ll see where all the Bonds years lie.

Ok, here’s the report for 2002.  I looked at each of the 28 possible off-cycles (days 1-7, 2-8, 3-9, etc) for each player.  I selected based on at least 75 PA in the off-cycle week and 150 PA in the three-cycle on period.  That left me with 317 hitters.

The mean difference was 87 points, with 1 SD = 38 points.  So, we expect, by pure luck, that 95% of the players to have a mean difference of around 10 to 160 points!

In the sample, Bonds was 36th out of 317 hitters.  #1 was this guy:
http://www.baseball-reference.com/m/mageewe01.shtml

He had a wOBA of .149 in his off cycle, and .347 in his on, for a difference of 199 points.

The #2 player is Jim Thome, on days 15-21 in the cycle (where day 1 is his first game of the season).  A difference of 182 points.

Palmeiro was 66th. Lawton was 29th.  The Giambis were 156 (the bad one) and 252 (the good one).  Sosa was 247th.

There *might* be something there.  But, there’s tons of noise, and whatever it is that you think you are seeing, it’s not that apparent.

I’ll run 1999-2001, and see if there’s anything else there.

Peter’s point is good, as we should look at pre-PED years.  1987 would be a good one, since the SD of wOBA is sensitive to HR, and 1987 offers us the closest to our current time period.

The difference I saw in DSG’s description was this sentence: “Starting with the first day of each season in which Bonds used steroids according to Game of Shadows.” If they assume he’s on for 21 days after that date, it severely constrains the data-mining.  But if just means they started the data-mining from that date forward, doesn’t matter.

* *

As Rally/26 indicates, shouldn’t we be looking at Bonds’ differential relative to his performance level?  If we use On-off-gap/season-wOBA, and Bonds is something like .130/.500 on average, that would be .26.  That would certainly move him much closer to a typical performance gap.

Ok, I re-ran for 1999-2001, in addition to 2002.  I have 1274 players in my sample (a little over 300 players per year).  Where did Bonds end up in terms of difference between off/on cycles?  58th, 81st, 124th and 213th.  Essentially, all at the +1 SD and higher.

What if I merge all the 4 years totals?  If I put in a requirement of at least 400 PA in the off cycle, I get 174 hitters in my sample.  The #1 guy, with an average difference of 140 points is… Barry Bonds.

(If I lowered the threshhold to 300 PA, Geoff Blum would be #1.)

David Justice (135 points) is #2.  I-Rod was #5 (123 points).

Palmeiro on the other hand was #69 out of 174 (average).  Matt Lawton was #95.  Jason Giambi was #102.

The guy at the bottom of the list, the one who showed the least amount of difference in performance between the 7-day “off” cycle and 21-day “on” cycle was Torii Hunter.

So, where does the leave us?  First, the idea that Derek had was tremendously insightful.  The execution however (looking only at 2002) was poor.  As I’ve shown, if you look at all players in 2002, Bonds doesn’t really standout much.  However, if you take 1999-2002, he stands out.

We should also remember that I’m taking the wOBA *differential*.  Bonds is starting from a much higher peak here.  If I *divide* his off cycle wOBA by his on cycle wOBA, he comes in at #34 out of 174.

Marquis Grissom (Brewers, Dodgers from 1999-2002) now becomes #1 and David Justice is #2.  Luis Gonzalez is now the guy who shows the least amount of difference in performance.

I think there may be something there, but alot more work needs to be done.

I’ll publish my data, so you guys can have a crack at it.

Here’s the Google Docs

The first tab is TOTAL, which is the aggregate of everything in tab SEASONAL.

To read this line:
batterid wOBA_off PA_off wOBA_on PA_on diff ratio
bondb001 0.392 528 0.532 1656 (0.140) 0.737

Barry Bonds had a wOBA in his off-cycle of .392, on 528 PA (excludes IBB, bunts), and .532 in his on-cycle on 1656 PA.  The difference is .140 wOBA, and the ratio of off-cycle to on-cycle is .737.

If you click on the Seasonal tab, there is also a column called “cycle”.  The cycle goes from 0 to 27.  The first line shows that Ivan Rodriguez’ worst 7-day off-performance, relative to his 21-day on-performance took place in cycle 17, meaning days 18-24, then 46-52, then 74-80, etc.

Day 1 is the player’s first game of the season.

IRod’s 2000 season is the largest outlier in the dataset.

***

You’ll also note *positive* numbers at the bottom of the sheet.  This is kind of peculiar, since I looked for the worst performance.  Theoretically, it should all be negative.  However, if you pay closer attention, you’ll see all these players have low PA totals, near the 75 PA threshhold.  Since I needed a minimum of 75 PA, it’s likely that of the 28 possible cycles that I tested for each player, only one cycle met the threshhold.  I guess I should have put in an additional criteria to have at least 5 possible cycles to choose from.  These guys don’t really tell us what we want.

I suggest that you add an additional criteria of at least 225 PA in the on-cycle.  This should solve this problem.

Updated Google Docs

Ok, this one has the SEASONAL sheet with 75 PA threshhold for the off, and 225 for the on cycles.

The TOTAL sheet is the aggregate of the SEASONAL sheet.

I suggest you use this file, and not the previous one.

Note that the average difference is around 90 points, with 1 SD = 30 points.  So, we should find 95% of players to be between 30 points and 150 points.  (Forget what I said about this part in post 33, since that relied on the larger dataset, with the bad threshholds for the on cycle.)

I haven’t figured out what pure randomness would expect.  Hopefully, someone here can pick up the slack.

To me, the ratio is the key variable here.  If we look at that, and look only at players with more than 1000 PA (n=245), Bonds ranks 72nd out of 245.  The mean ratio is .762, SD is .044, Bonds is .737, so he’s 1 SD above average (low ratio = big disparity).  Doesn’t impress me.

Among the 25 most consistent hitters by this measure are: Giambi, Bagwell, Helton, and L. Gonzalez (#1).  If suspected PED users are clustered at one end or the other—I’m not sure they are—it’s probably on the most consistent end of the scale. 

If we limit the list to more similar players, there are 106 guys with 1500+ PA and .350+ wOBA in Tango’s sample.  Bonds ranks 19th out of 106 in terms of having smallest ratio.  Yes, his disparity is on the high end, but not off the charts.

Another one of the consistent players is Ken Caminiti.  Its pretty clear to me that this method is not going to identify steroid users.

Derek seems to attract the “bad” crowd:
http://www.amazon.com/gp/product/customer-reviews/0618551131/ref=cm_rev_sort/102-2368645-3089749?customer-reviews.sort_by=+OverallRating&s=books&x=7&y=9

Lots of good reviews, but a couple of bad ones.  That “MantoFan” not only posts his review, but also puts his comments on at least two other reviews.  At least Derek has lots of good fans who can drown out the others.

Phil also posted his review here:
http://sabermetricresearch.blogspot.com/2007/07/cheaters-guide-to-baseball-and.html