THE BOOK--Playing The Percentages In Baseball

From the article:

Even with that though, the players didn’t do all that well in terms of UZR, with an average of -.28 per 150 games played.

Didn’t do all that well?  Moving from another position after the age of 30? -.28 UZR? Essentially an average CF. That is darn good to me…

I really think it would help alot if UZR is never ever quoted to anything other than an integer.

It would also help if UZR were quoted with some sort of margin of error. I know, it’s supposed to be a measurement, not a sample, but we know for sure there’s a lot of uncertainty attached to it.

My question on reading the article: is center really that much riskier than RF or LF? If Werth doesn’t try to imitate Ken Griffey, Jr. I’d be surprised if it mattered much.

2/Tango - agree, I publish fielding runs to 1 decimal, but in conversation always use integers, even sometimes rounding to nearest 5

This is something I started looking at as a result of Cabrera being moved back to 3B. I developed an algorithm to find the optimal arrangement of players, given their batting and stealing runs above average, and their fielding runs above average at each qualified position (even with giving Cabrera -12 at 3B, putting him there does maximize the lineup runs above average).

I haven’t figured out how to code this yet, but did dump the offensive and defensive numbers into a spreadsheet and was playing with it manually to figure out the correct methodology.

The questions that come up include what ‘true talent’ fielding runs above average per 162 to give to someone who has only played a few games (or none recently) at a position. It’s fine to say that that someone actually had, in the past, a rating of +2 in a limited number of games. But if I’m going to do a per162 rating, I’ll want to regress it. Normally we regress to league average, but if I player has only minimal (or no) playing time at a position, that might be information telling us that the coaching staff doesn’t think he’s capable, therefor if played there long term he’d likely be bad. I plan on doing a study of comparing different sets of defensive data, such as regulars at one position doing occasional time at another, or position switches, to see how the positions relate. Also along that line, if I have a good sample for one outfield position, that should give me some information of how the player would do at another outfield position. Like in Strat-o-Matic,, a get’s a 4 rating for positions with small samples. In my case, I want to discourage my lineup generator from putting a player at an unrealistic position - he may have played some there in the past, but that doesn’t mean he can handle it full time, even without bad results so far.

Back to Werth, over the past 7 seasons, from 2005-2011, he has started 92 games in CF in MLB, 19 in 2010 and 18 in 2011, so he’s not a total stranger there,

For the 7 year period, MLB only, I have

Pos  GS  Inn TM_BIP AirOpp GBOpp AirOut% FRAA per162 
 LF  51  768   2362   296   53   .497   -9.7  -2.2
 CF  92  841   2533   416   60   .623   +6.6  +2.2
 RF 598 4843  14582  1983  314   .570   +5.3  +3.4

For the per162, it’s a regressed projection from the last three years, I haven’t attempted yet to blend the ratings from the similar positions.

Would something like the following make sense?  Regress his CF +/- based on his RF +/- minus the average difference Tango found in getting the position adjustment for WAR.  I don’t know how to get the playing time of the regression factor.  So, if he was +5 in RF we take his limited CF +/- and add maybe -5 if the position difference is 10/150 times some number of games (300?, wag) apply an age adjustment and we have our regressed estimate.

I need to do all the matchups in the data, but I would think of doing it like splits - look at the ratio between the two, and regress the observed split (primary position to secondary one) to the standard split. Then apply the resulting regressed split ratio to the two combined to get the rating for each.

(per162+pos adjustment)*IP for all three then divided that number by total innings puts him at -3.8 per162 for his OF career. Back the positional adjustment out and he’d be -6.3 in CF and + 3.6 on the corners. Looks to pass the smell test based on those numbers.

^^^^Based on the numbers in #4.

Brian, #6, that assumes that player “splits” between positions are real. They probably are, however, without doing some research, we have no idea of the spread of “positional split” talent is, so you don’t know how much to regress.

I simply combine a player’s UZR at all positions after adjusting each one using “Tango’s positional adjustments.” Then I can simply reverse adjust that position-neutral number to any position I want.

So if Werth is +3 per 150 in RF and LF for 500 games and +1 per 150 in CF for 100 games (made up numbers), I simply adjust and combine, so, if CF is 5 runs per 150 “harder” than RF and LF, we have an “adjusted” (to CF) UZR of -2 in RF and LF. Combine that with the +1, and we get -2*500 + 1*100 / 600 or -1.5 per 150 in CF.

I like the idea of regressing a player’s own splits to the standard ones (the 5 run difference between CF and RF/LF), but, again, I have no idea how much to regress.

I think that the regressions would be quite different for different positions. For example, I doubt that there is much of a spread in “split” talent for players moving around in the OF.  For players moving around in the IF, there probably is some.  For players going from OF to IF or vice versa, there is probably a lot of split talent…

mgl, I’ll look into applying Tango’s positional adjustments as something I can do in the short term, but I plan on doing the research you describe.

Tango’s positional adjustments are based on players who switched positions, so the numbers are pretty solid. Now all you need to do is figure out the spread of talent for the splits. You can do an intra-class or year-to-year (or odd/even games, etc.) correlation or something like that. 

If you do that let us know what you come up with!

BTW, do you do that with home/way splits for players, both for park effects and HFA? Do you regress their own splits toward the league norms? If yes, how do you know the regression amounts for that? That is something I’ve been meaning to do…

We had a little thing about personal-park factors a month or so ago, discussing Bonds’ HR splits at 3com (or whatever the park was called ATT?).  And that while LHH had a -33% HR split at ATT, Bonds hit as many HR at home as he did on the road.

And we ended up concluding that the regression amount was about 3000 PA.  So, we’d have that Bonds was hurt by about -16% at home.

Something like that…

But, yeah, plenty of room for research.

11/mgl - right now I have park factors for right and left handed batters, but no further individualization. I did an article a few years ago at StatSpeak which showed that players with the highest hr% were least affected by the ballparks, likely that a higher pct of their homers were no doubters.

You have a good idea there that I hadn’t thought of.

What I am planning on looking into is dividing the outfield into seven slices, then using WOWY to see how each batter compare to his peers, holding for bat hand/ballpark/outfield slice.

"Didn’t do all that well?  Moving from another position after the age of 30? -.28 UZR? Essentially an average CF. That is darn good to me… “

Isn’t there some selection bias in it too? You aren’t moving a bad OF from the corners to CF. It’s not like you will see Manny Ramirez move to CF in his 30’s. So only the good OF (Ichiro, Werth) are moving to CF.

Preliminary results

I compared my FRAA calculations at the three outfield positions. All levels, same team each season and position, a three season time span.

So a LF in 2011 would be compared to how he died in CF and RF in 2009-2011, if it was the same team.

The FRAA calculations were done by WOWY which considered ballpark, but I decided to hold the team constant to avoid park factors as much as possible.

Run values are scaled to 700 touches, a full season for LF or RF (CF max is about 800)

Pos1 FRAA Pos2 FRAA  Diff   Size
 LF +4.71  CF -2.55  7.26 265306
 LF +0.32  RF -0.71  1.03 315452
 CF -2.50  RF +3.88  6.38 243135

Corner outfielders who also play CF are better than average (4 to 5 runs) in the corner, but are still 2 to 3 runs below average in CF. A 7 run difference between CF and a corner.

Fielders perform about a run better in LF than they do in RF. These are the same fielders, same team at each position. This is visible in not only the RF to LF comparison, but also in each of their comparisons to CF.

The way I have the matched pairs set up, I can vary the vary the levels considered (such as returning MLB only) and I can vary the number of years in the window (0 would mean both performances in same season).

Brian, does that include arm? Also, what fielding metric are you using again? IOW, where does FRAA come from?

Unless Tango has updates since then, in 2008, he had around 10 runs difference between CF and the corners.

Right, I used UZR from 2002-2010 or something like that.  It’s been pretty stable at 10 runs.

Considering how much difficulty teams have at finding CF, and how often a +10 and better corner OF does NOT move to CF, you gotta figure the higher figure might be better.

Like I said, preliminary, my first attempt and I appreciate the comments.

It’s my own calculation that I use as Oliver’s fielding runs. At this time it does not include outfield arm, but I do now (since mid 2011) have the sufficient detail of play by play to calculate. I just have to find the time write the code that gets the data into the warehouse and start analyzing it. But it does include extra bases gained by the batter on hits, which I believe UZR does not.

To the question of how good a corner outfielder do you have to be to play a credible center field? Using my own historical ratings, and assuming the same 7 run difference between left and right compared to center.

The #1 ranked qualifying CF from 2005-2011 had an average of +19 runs, ranging from seasonal highs of +12 to +26 (Gutierrez). The worst average -13, ranging from -6 to -22 (two worst Griffey). If an average corner OF (0) is -7 in CF, that would rank among qualifying CF each season from 26th to 31st, a mean of 28th.

The range in LF was a max of +10 to +21, mean +15, min of -10 to -19, mean -15. RF max +9 to -21, mean +13, min -11 to -25, mean -15. For a corner OF to be average in CF (0), he’d be +7 in a corner, which would rank from 4th to 7th among LF, mean 5th, 4th to 10 in RF, mean 7th.

So, to play an average CF, you’d need approx a top 6 corner OF. An average corner OF would be on avg the 28th best CF.

Werth has improved (in my ratings) each of the past two years. He’s been a qualifier (top 30 in touches) each of the last three seasons (2009-2011), ranking 22nd (-2), 15th (+2) and 4th (+7).

The 5/4/3 weighted mean of his last three seasons at all outfield positions is +4.2, with 12.3% of his touches in CF. Assuming a corner OF has a mean true talent of 0 in LF or RF, -7 in CF, take that -7 * .123, his expected would be -0.9, so I’ll give him a true talent of +5.1 in a corner, which would be -1.9 in CF.

It is possible that a fielder’s ratio of CF to a corner is better or worse than the mean, that there’s a distribution. We’d need to find how much regression is necessary. Three players made the top 30 playing time in each of LF and CF from 2005-2010. Damon was 4th in unadjusted catch rate in LF, 13th in CF; DeJesus 6th in LF, 15th in CF; Pierre 9th in LF, 24th in CF.

If a team is going to to take a 7 to 10 run loss on defense by moving a player to a harder position, they would need to make up the difference offensively. My depth charts show Werth and Harper as the Nats best two outfielders. Morse can play LF or 1B, with Bernadina the next OF. The optimal alignment involves Harper pushing LaRoche out of the lineup, and Morse’s glove out of the outfield.

If a team is going to to take a 7 to 10 run loss on defense by moving a player to a harder position, they would need to make up the difference offensively.

Brian/Tango:  Have either of you ever tried to estimate how many runs a team would actually lose by moving a player like this?  If we assume constant offensive peformance, and replacement level player in the other position, how many runs would the Nats lose by switching Werth to CF?  Maybe 2 runs?  Less?