THE BOOK--Playing The Percentages In Baseball

Do you have a link to the original article?  Thanks.

The original article:
http://www.beyondtheboxscore.com/2009/8/10/983766/graph-of-the-day-raul-ibanez

Thanks, fixed.

I think this does a good job of showing a concrete example of how the raw numbers are inflated from one park to the other, which is important to note when you look at Ibanez’s line and see the big improvements.  When it is translated to a win value, though, doesn’t the 10-run per win estimate obscure things because each park here has a pretty different scoring environment?  If the estimate is that Ibanez picked up 5 runs of production going from Safeco to CBP, I think that’s fairly close to the difference the park factors make, so a 5 run difference in such different parks shouldn’t translate to a half a win.  It’s pretty close to the same production in wins, just in a higher scoring environment inflating the run values.  Unless I’m missing something about how these translations work (fully possible), it seems like these numbers are more just reinforcing the magnitude of difference between the parks shown by park factors with a concrete example than they are showing that Ibanez’ win value is artificially inflated by CBP.

Nonetheless, I really like the graphs, and the effort to putting those together gives us some good information about just how the change in parks is affecting Ibanez.

I would be very careful with this kind of stuff.  Greg did the same thing with Adam Dunn (and Manny I think) in the Hardball Times annual.

The first person I saw make us of this was, unsurprisingly, MGL, several years ago, in what he called “virtual HR”.  He did exactly what Greg and the others have done, and plotted the flyballs hit, and figured how many would be HR in various parks.

And then he looked at players who switched parks.  And guess what: the actual HR hit were halfway between what would be expected from the virtual HR mappings, and what he actually hit the year before.  That is, there is a quality to hitting HR beyond just hitting it far.

Basically, what everyone here is doing is (a more elaborate version of) half of what MGL did.  What’s missing is the actual other half: the testing of the virtual HR.

What I was trying to convey was that Ibanez’s increased production isn’t necessarily real from a true talent standpoint.  What would be a better word to use here?  “It is important to remember not all of Ibanez’s increased production is sustainable?” I certainly see where you’re coming from here - the HR’s happened, and we can’t take that away from Raul.

"It is important to remember that some of Ibanez’s increased production is disproportionately affected by hitting at CFB and not hitting at Safeco”.  This is your thesis.

The thesis, however, relies on the fact that those HR he hit could be “mapped” to each ballpark.  This has not yet been proven to be true.  Indeed, there is some evidence that only half of that can be done.

Tango, I’m not sure exactly what you mean when you say “testing the virtual HR”, or mapping HR to each park.  Can you elaborate?

Going on memory here, but in last year’s annual, you showed how many HR Dunn and Manny would hit in each ballpark based on the dimensions of the park and the distance that flyballs travelled.

This is theoretical and based on the assumption that if Dunn hit his spray of flyballs in one year with Cincy as his home park, that he’d have the same spray with Washington as his home park.

But, is this true?

Could you, for instance, look at all players who have a new home park, and see if this is true?  That is, you are positing that the “virtual HR” is a better indicator of his true HR power than his actual HR (park-dependent) are.

Is it possible, though, that there is more to hitting HR than knowing the flyball distance?  That is, the best estimate of hitting a HR is based on both the flyball distance and whether that ball actually left the park (regardless of which park it left it out of).

Tom, are you referring to the article Greg published at Baseball Analysts:

http://baseballanalysts.com/archives/2009/02/2009_projection.php

I can’t answer that yet, as that site is blocked at the office.

Here are Greg’s main points in his outline of how his projection system works:

How It Works: Steps in the Hit Tracker Forecasting Method

* Observe all long fly balls hit by a player in the past 1-3 years.
* Analyze each long fly ball in its actual weather conditions, to determine its launch characteristics (Speed Off Bat, Horizontal and Vertical Launch Angles, Spin).
* Note each long fly ball’s original result (2B, 3B, HR, Flyout, etc.).
* Project each long fly ball into each of the 30 MLB ballparks, in the average weather conditions for that ballpark (calculated over a 5-year period).
* Note the hypothetical result of each projected fly ball in each ballpark.
* For each ballpark, count the net hits and bases for the long fly ball data set:
* Apply the net adjustments to hits and bases for all the long flies to the player’s actual stats for the season in question. Calculate OBP/SLG with the adjustments. This becomes the player’s projection for that ballpark.
* For projections based on multiple years of long fly balls, apply appropriate weighting factors (e.g. 3-2-1) to the projections for each ballpark.
* Using the MLB schedule for the season of the projection, create a projection for the player as a member of each team by multiplying their performance averages in each ballpark by a weighting factor proportional to the number of games each team plays in each park.

He applied this in detail to Manny Ramirez and briefly to a handful of other players.

It is certainly possible that another factor such as the one you propose is lurking, but so far, the HR projections are doing pretty well for the guys I did 3 year projections on:

Ramirez: projected 36 in 143 games, actual 12 in 59 games.  36/143 projects to 15 in 59, so I am low by 3 here.

Dunn: projected 43 in 156 games, actual 30 in 111 games.  43/156 projects to 31 in 111, so low by 1 here.

Bay:  projected 27 in 152 games, actual 21 in 103 games.  27/152 projects to 18 in 103, so high by 3.

Collectively, projected HR’s by this group of players = 64, actual so far = 63.  Seems like the system is on target, and showing some of that random variation you just can’t get rid of when the outcomes are binomial-based (HR or not HR).

Now, the guys I did 1 year projections on are showing a lot more variation, but not too bad overall on the total: Teixeira is +6, McLouth -4, Holliday -6; the projected total for those guys together was 63, and they have totalled 59.  Consistent with the concept that three-year projections ought to be better than one-year projections…

The baseline would be Marcel:

Manny 27 in 584, actual 12 in 248, or 11.5.  Off by 0.5

Dunn 32 in 589, actual 30 in 470, or 25.5.  Off by 4.5.

Bay 24 in 596, actual 21 in 445, or 18.  Off by 3.

Yours was a total absolute difference of 7, and Marcel is 8.  I definitely don’t see it as a win for either side.

It is interesting that the total for you came out close to even, and Marcel ended up low.  This is to be expected when players are having good seasons (since regressions presumes that none are as good as they’ve shown, but these three guys are showing that).

As I said, the best way to test is to look at a large number of hitters who switched home parks, and then baseline against Marcel.

I have to believe that the “virtual HR” method only gets you there about as much as Marcel does.

Tom, let me also bring in OPS for these guys - because the method I came up with not only projected HR’s but all batted balls that landed within close proximity to the shortest fence in any of the 30 parks (this is I was sure I got all the potential HR balls).

OPS

Name HT Marcel Actual through 8/9
Dunn .944 .871 .984
Ramirez 1.071 .923 1.013
Bay .869 .833 .873

If you work that out, my system has an average absolute error of .034, while Marcel has .081

Now, of course these are just three players, well-established ones at that, but that seems like a bigger difference.

It’s narrowed a lot in the past few weeks, also.  As of 7/19, the AAE for HT was 0.014 to Marcel’s 0.099 (/cherrypick)

Great!

It’s this stuff that should be done for all the forecasting systems.

I am disappointed that consumers don’t demand these kinds of validations.  But then, everyone who sells a forecasting system will be exposed much more than they’d want to be.