THE BOOK--Playing The Percentages In Baseball

Well, Clay nailed the spread.  One SD is in fact 9.0, so great job on Clay there.

Beat me to it.  I wanted to calculate it so that I could ask you why it’s important for determining whether it’s a good prediction or not.  Has the SD in wins proven to be remarkably consistent around nine over the years?

Does anybody have an idea as to how Clay comes up with his projected win totals and run totals?  There is no word about method at the site linked to.  I’m just curious.

The biggest outliers I see imho are…

AL East
BOS-98, NYY-92.  I would put the Yankees a few wins ahead of the Red Sox with neither of them at the 98 win level.

NL East
I think the Miami (87) numbers is a little high.

AL Central
I think the Tigers (85) margin over the White Sox (84) should be higher, otherwise the division looks ok.

NL Central
CIN(88), MIL(87), STL(83)… I think the Cardinals should be at the top 90+.

AL West
TEX(102) is too high.  I’d probably have them with the most wins in the AL but more like 95-96 ish.  The Athletics (80) is too high.

NL West
This division is such a crapshoot.  I’d probably give the Giants (83) a few more wins though.

Clay runs one million sims…

True talent is spread at one sd = .060 wins per game.

Our ability to estimate that must be lower than that (since we are not god).

You can run a correlation of Rally and MGL and other’s forecasts through the years, and best-fit it to the actual results, to see what would have been the ideal SD.

Looking at the Red Sox numbers, Clay has them with around 58 WARP. So that places his replacement level at 40-122; we know Clay likes a lower replacement level than most.

Moreover, he has too many PAs for Crawford, Matsuzaka (injuries) and Scutaro (traded today). Beckett being projected as highly as Lester also strikes me as incorrect since if you do a simple weighted average of the last three years, Lester comes out way ahead. At that, nothing else strikes me as out of line with the individual projections.

What kind of simulations?  Does he have a simulator where the teams actually play each other using the actual 2012 schedule (1 millions times)?  Or some other kind of model?  Thanks!
vr, Xeifrank

The RAR numbers for pitchers look off, or maybe I just don’t understand how Clay’s calculating them.  It’s definitely not based on projected runs allowed unless the park factors for players are different for guys on the same team. 

Example:

Hisashi Iwakuma is projected to allow 87 runs and pitch 173 innings, which translates to 15.7 RAR, or 0.82 RAR/9. 

Blake Beavan is projected to allow 89 runs and pitch 165 innings, which translates to 21.2 RAR, or 1.16 RAR/9. 

In this case, fewer projected innings pitched and more runs allowed equals a higher RAR.  Since we’re looking at two same-handed pitchers on the same team, I wouldn’t think we’re dealing with a park adjustment issue.  Both guys are projected to make all of their appearances as a starter, so it’s not an issue with a different replacement level for relievers. 

If anyone’s seen Clay’s explanation for how he does pitcher RAR, I’d love to look at it.  Just based on a cursory look at the results, some of the data looks a little odd from here.

Maybe one SD should be 8… I’ll have to work it out.  I definitely wouldn’t expect to see it less than 7 though.

I find it interesting that the Yankees and Red Sox have a spread of 0.67 StDevs and the Rangers and Angels have a spread of 1.56 StDevs. Otherwise, none of the results appear out of whack.

Red Sox: Pitching RAR = 269
Hitting WARP = 33
Yankees: Pitching RAR = 265
Hitting WARP = 34

Texas: Pitching RAR = 317
Hitting WARP = 31
Angels: Pitching RAR = 268
Hitting WARP = 29

I also find it interesting that Yu Darvish is projected +71 RAR, while Sabathia (+54), Lester (+55), and Weaver (+60) are much lower.

Texas definitely looks to grade out higher than the Angels, but probably not 14 wins worth, while the Red Sox and Yankees seem about even.

It looks like Clay is using standard values for playing time of various positions.

For example, on the Phillies:
* Howard (starting season on DL) is given 138 games played, more than Utley.
* Brown (questionable fielder, probably in the minors unless brought up by injuries) is given time in CF, Mayberry (presumably the primary CF backup) isn’t.
* Four players are given 16 games of “DH” time; Thome doesn’t get any more than anyone else.
* Valdez gets more time at 2B and 3B than SS.  With any reasonable backup 3B, Valdez shouldn’t get any time there.

Clay also has the average AL team at 82 wins and the NL at 80 wins.

Is there a Marcel for team standings?  I suppose it wouldn’t be hard to convert forecasted ERA and wOBA into WAR, and then sum the WAR per team and add it to the baseline for whatever replacement level is used (a .300 team I believe for Marcel).

I’m not sure whether that would still be simple enough to bear the name Marcel though..

There’s no reason to convert to WAR.  All you want is runs.  You have runs allowed and runs for.  Convert that to wins.

I’m with Dave Cameron on the RAR thing too. Look at Daniel Bard and Clay Buchholz’s projections. Bard Allows 20 less runs in only 9 less innings, yet has 2 less RAR. Although Dave’s example probably is better since Clay has Bard with 19 relief appearances.

Xeifrank and Phil D, why are the Yankees better than the Red Sox? Using Clay’s projections and Tango’s markov simulator, I have the Sox at 838 runs and the Yankees at 808. Then Clay has the sox at 666 runs allowed and the Yankees at 701. Using Pythag that’s 91, almost 92 wins for the Yanks, 98 for the Sox.

#16. Because that is what my simulation results are showing.  Also, Vegas futures have Yankees at 5:1 and Red Sox at 8:1.

I have Boston at (839, 760) and Yankees at (819, 700).  So I have some big differences with Clay concerning Red Sox run prevention. 

I am not saying my numbers are perfect, they likely aren’t.  But mine match better with the Vegas future odds, that’s for sure.

Xei: can you post

1. SD of wins for Vegas
2. average wins for AL, NL

#18.  I haven’t seen that data yet.  I will ask around if nobody else has it.  All I’ve seen so far are “odds” on winning W.S.  One SD for my sim is 7.98 though.
vr, Xei

what I don’t get is how the Yankees have better run prevention than the Sox. The Red Sox had Buchholz injured for almost all last season, and bard is moving to the rotation. Whoever is in the 4/5 starter slots is almost sure to be better than Lackey and Weiland/Miller/Wakefield

At that, nothing else strikes me as out of line with the individual projections.

The relief to starter projections seem a little too optimistic to me.  Just eyeballing them, he seems to have Chris Sale, Neftali Feliz, Josh Bard and Alfredo Aceves projected too favorably if they’re going to be starters.

My standings projections look nothing like this.  I’ve got:

NYA: 95.9
TEX: 92.9
BOS: 92.4
PHI: 92.1
TAM: 91.8
LAA: 91.5
STL: 88.9
MIL: 87.3
CIN: 86.9
DET: 85.7
ATL: 85.6
ARI: 85.4
SF: 84.6
CLE: 83.9
FLA: 82.2
WAS: 81.5
COL: 80.8
OAK: 78.7
TOR: 77.0
LAN: 75.9
SD: 75.4
CHA: 75.0
NYN: 75.0
KC: 73.4
SEA: 72.9
CHN: 71.2
BAL: 69.3
PIT: 68.6
MIN: 67.0
HOU: 61.2

Standard deviation is 9 on the button and AL average is 82 wins, NL average is at 80.

The numbers I have on hand are from 2001-9.

Standard deviation for MLB varied from 9.3 wins (2007) to 14.8 wins (2002), averaging 11.9 wins.

NL average wins varied from 79.2 (2006) to 81.6 (2003), averaging 80.4 wins.

AL average wins varied from 80.2 (2003) to 83.0 (2006), averaging 81.6 wins.

Standard deviations in the NL varied from 8.0 wins (2007) to 14.1 wins (2004), averaging 10.6 wins.

Standard deviations in the AL varied from 10.9 wins (2007) to 17.1 wins (2002), averaging 13.5 wins.

Here are the win total over/under numbers from the previous season.

http://www.vegasinsider.com/mlb/story.cfm/story/1180098

Here are the SD calculations on those numbers.
NL: 7.78
AL: 8.03
Overall: 7.76

And the average wins by league shake out to be…

NL: 81.09
AL: 81.57
Overall: 81.32

Your guess is as good as mine as to why the averages are slightly over 81.  Could be that more people like to bet the “over” (think fans betting on their teams to do well) so they bump the numbers up ever so slightly.  Could also be some rounding error in there.
vr, Xei

By the way, true-talent variation should not be as high as the observed variation.

Expected true talent variation can be found by deconvoluting with the Gaussian error: (Talent)^2 + (Luck)^2 = (Results)^2.  The “luck” term is 6.364 (half the square root of 162).

So a real-world standard deviation of 11.9 wins (ML average over 2001-9) ought to correspond to something more like a 10.1 win standard deviation in true talent.

Jeremy is correct.  I always use one SD = .072 wins per game for observed, and .060 for true.

Since we are not god, we cannot estimate a true of 1 sd = .060.  Therefore, I figured .050 would make more sense.  But then there’s also the issue of playing time not being fixed/known, so one SD = .045 probably seems more reasonable.

.045 x 162 = 7.3

So, that’s probably around where you want your estimate to be.

For the people running the sims: you are almost certainly using the true talent values with uncertainty of zero.  A good sim will vary the true talent level after every run.