Tangotiger Blog

Wednesday, August 28, 2013

Uncertainty in offensive metrics

Colin takes a stab at explaining the uncertainty in offensive metrics.

The "uncertainty" that Colin is reporting is the difference between the standard linear weights run values (the 1.4 for HR, the .3 for walk, etc), and the run values based on the 24 base-out matrix (the example he gave for the bases loaded strikeout, etc).

Fangraphs tracks both values (see wRAA and RE24, though I'm not sure if there are park adjustments to contend with), as does Baseball Reference.

So, what Colin is saying is that there's a 90% chance that Trout's RE24 in 2012 is going to be higher than Cabrera's RE24, if we rely on their standard linear weights values as being 12.5 runs apart.?

Sunday, August 25, 2013

oWAR + dWAR does not equal WAR

?The correct way to handle WAR within the WAR framework is that each component is compared to the average, and then you have a replacement level component (tied to playing time) at the PLAYER level.  This is the right way to do it.  Not "a" way, but the right way (if you subscribe to the WAR framework).  Fangraphs presents it as such at the bottom of each of their player pages, as does Rally on his site.

If you want to start to merge components into "offense" and "defense", you still need a THIRD component (the replacement level or playing time).  So, it's offense relative to average, plus defense relative to average, plus the replacement level (average relative to replacement level).  Andrelton Simmons, using rWAR, breaks down as follows for his career:

-10 runs offense relative to average

+65 runs defense relative to average (which includes his positional adjustment)

+22 runs average over replacement

===

+77 runs above replacement

In the BR.com presentation, it references "oWAR", which you might think is "offense above replacement level", when the reality is that THERE IS NO SUCH THING as replacement level offense.  It's simply something that we must get away from talking, referencing, or otherwise whatever.  It's anathema to the WAR framework.  What "oWAR" should be is the "offensive component of the WAR framework".

But like I said, at its barest, the WAR framework has THREE components.

***

Now, if we must and insist that we only have two components (offense and defense), you then have to decide how to distribute the third component, the playing time one.  There is zero reason to simply distribute it all into the offense component. 

***

All of the above is all about the WAR framework, and it's not up for discussion.  What is an opinion is what is about to follow.  If we insist on distributing the playing time portion onto the offense and defense, how to do it.

So, one idea I have is that since the spread in offense is about 1.7 times the spread in defense, then we'd split the playing time component accordingly.  So in Smmons' case, the +22 runs will get spread at +14 for offense and +8 for defense.

Therefore, oWAR would be +4 runs and his dWAR would be +73, and that gives him his WAR of +77.

Friday, August 23, 2013

Explain FRA to me

I have nothing good to say about Fair RA.  Momma told me I should stop there, but, who listens to their mother really?

***

While the idea behind it seems reasonable, the results are frankly ludicrous (see the whole thread, but especially the comments starting at comment 21 if you are short on time).? 

So in the end, I can't explain the methodology nor the results.  And yet FRA is the central component to WARP.  Outside of Colin, I doubt there's a single person at BPro that can explain FRA that can be understood by any one of us. (That's a challenge.  Crickets will be here until it's taken on.)  But it is being referenced by its authors, most visibly in their book (where they highlight the metric by showing that Doc Halladay wasn't as good a pitcher as the rest of us think). 

The thing is that FRA has been around for three years now (Happy belated Birthday!), and I haven't seen any effort at all in moving the understandinging of the metric forward.  It simply exists, and it is simply being referenced.  I think outside of the original article, no article has ever been devoted to it since.  The sellers can't sell the metric, and the buyers aren't buying the metric.  So, why does it even exist?

SIERA in contrast made a splash at BPro when it came out, and Matt had several articles that delved into SIERA and what made it tick.  And almost as quickly, it was gone, in a takedown by Colin himself.  (SIERA survives at Fangraphs.) 

The last time I took out the daggers was to eviscerate the PECOTA percentiles.  They were wrong from the start, and they were wrong for several years.  And yet, there they stood, existing without explanation.  Until Colin proved that they were wrong.

When will the FRA obituary be finally written?   Until then, someone, anyone, explain FRA to me, in a way that I can explain it to someone else.

Wednesday, August 21, 2013

Reworking WARP

WARP is the BPro version of WAR, which in my nomenclature would be called pWAR. Colin says he's going to rework it. It'll be interesting to see how different it'll be from the framework I developed, or from the implementations (rWAR, fWAR) of that framework.

Colin is right that the place where it's going to get interesting is the pitcher-fielder interactions.  Currently, WARP is based on FRA.  FRA might be rooted in logic, but there's an obvious problem, given the results we see of Felix, Doc, and Maddux.  That makes FRA something that I discard (see this thread from last year, and if you are pressed for time, start at comment 21).  Anyway, I presume we will get drastic new results for these three pitchers among others, and they will fall in-line somewhere around fWAR and rWAR.

Thursday, August 01, 2013

Standard wOBA

OBP's formula doesn't change, regardless as to the relative value of the HR to the ngle, given the circumstances.

SLG's formula doesn't change, regardless as to the relative value of the HR to the single, given the circumstances.

The core of FIP doesn't change, even though it should.  (It does have a fudge factor though.)

It seems to me that what helps these metrics is that they have a standard formula, and then analysts can tweak them to suit their purposes.

wOBA should be like that, have a standard formula.  This is what I use:

Version 1 - Basic

numerator:

0.7: UBB+HB

0.9: 1B+ROE

1.3: 2B+3B

2.0: HR

denominator: ?PA - IBB - SH

 

Version 2 - Speed

numerator:

0.7: UBB+HB

0.9: 1B+ROE

1.25: 2B

1.60: 3B

2.0: HR

0.25: SB

-0.5: CS

denominator: ?PA - IBB - SH

If you don't have ROE handy, then ignore it.  If you don't have walks split between UBB and IBB, then ignore the distinction.  If you don't have SH, ignore that too.  Same deal with steals.

So, that's what I'd like to see for non-Fangraphs sites.  It's nice and simple.

The key point by the way is that you start with "1" for any positive event (i.e., the numerator of OBP), but adjust it up and down relative to the event.  The overall average of these events will come out to "1", which is why it is called a "weighted" on base average.  Both metrics (unofficial OBP and wOBA) share the same denominator.

Monday, July 29, 2013

wOBA v Runs / PA, 2003-2012, Team Offense

?Data from Fangraphs.  There's 30 data points, one for each team.  That means ten years of data, per team.

And this one is per team-season (300 team-seasons in all).

Thursday, July 25, 2013

Linear Weights by the 24 base-out states

?Matt continues his look at RE24, but he's trying to factor out the baserunning part of RE24, when looking at the change in base-out states by batter.  In other words, instead of the generic Linear Weights metric, he's doing a separate Linear Weights for each of the 24 base-out states.  Something that looks like this:

http://tangotiger.net/RE9902event.html

Or this:

http://www.tangotiger.net/lwtsrobo.html

We have a pretty long description of it in The Book, at least as it pertains to the HR.

Now, Matt talks about the difference between the actual RE24 and a Linear Weights approach by the 24 base-out state as "baserunning".  But, there are other reasons.  If a hitter say hits alot of "sliding doubles" (singles that they stretch into doubles), then those doubles will only move baserunners over by two bases all the time, and rarely three.  Similarly, with infield singles, those would rarely, if ever, move a runner from first to third.  

So, a double is not a double.  And a single is not a single.  You'd have to look at where those balls ended up.  That said, the differences that Matt shows makes it seem that this is not really that big an issue.

Thursday, July 18, 2013

wOBA, wOBA, wOBA

Yowza on the comments following this article discussing wOBA.  Just picking out a couple while I have a minute here:

Why do you think sites like baseball reference do not show wOBA. Is it proprietary to fangraphs or is there more to it.

There's nothing proprietary about wOBA (or FIP for that matter). 

Why is the coefficient for a HBP not the same as a walk. They are essentially the same thing no?

Hit batters occur more randomly.  Walks occur disproportionately with first base open.  So, overall, the hit batter generates more runs, just as a regular walk generates more runs than an intentional walk.

Why are productive outs penalized? If there is a runner on 3B with less than 2 outs, depending on the game situation, a SF is better than a walk or HBP.

If we were to record ALL outs, then I would treat them differently.  For those people, we have RE24, which looks at each event, and records what ACTUALLY happened.  A strikeout with man on 3B and less than 2 outs is a huge cost to the batter, and this is recorded by RE24.  A strikeout with bases empty is treated identical to any other out.  A sac fly is given its just due (moves runner from third, and batter records an out). RE24 is perhaps the greatest thing sabermetrics can offer the non-saber follower.  It's the ultimate bridge stat that can be equally useful to all comers.  I'd like to see articles every week on RE24.

I agree that IBB and sac bunts shouldn’t be included, though.

More often than not, I set IBB and sac bunts aside in virtually everything I do, be it OBP, wOBA or whatnot.

***

Anyway, I skimmed over most of the comments, because there was too much back-and-forth banter to pick out any questions.  If there are other questions, please post them below.

"If you wOBA me, I will wOBA you."? (Youtube)

Saturday, July 13, 2013

OPS Primer

?Patriot does a good job in getting into some of the nuts and bolts of OPS and OPS+, their biases, and how it relates to run scoring.

Monday, June 03, 2013

Weighting of OBP relative to SLG

Phil has devoted several threads to correlation of OBP, SLG to runs scored.  He uses actual data.

In this post, I'm using a very controlled set of data, with the Markov calculator, which is a perfect mathematical calculator.?  If you can get past the idea that I'm not modeling outs on base, we know precisely how many runs will score, given the frequency of events, and the transition rates off those events.  If you aren't using the above calculator, you are missing out.  Use that calculator.

Anyway, I created three different leagues.  One league is the standard league (4.9 runs per game), another is a low scoring environment (about half the standard) and the other is a high scoring environment (double the standard).

For each league, I created nine different hitting lines.  I did this by adding 0.1 walks or adding 0.1 AB and 0.1 hits, and so on.  The result of the correlation follows.  For the standard league:

Runs per game = 22.2*OBP + 11.7*SLG - 7.4

But, we're really interested inthe relationship between OBP and SLG.  And 22.2/11.7 is 1.90.  That is, you want to weight the OBP at 1.90 for every 1.00 of SLG.

For the high-scoring league, the correlation has the weight of OBP/SLG at 4.1.  That is, we REALLY REALLY want high OBP players, if the run enviroment is high.

For the low-scoring league, the correlation has the weight at 0.84.  In this case, since the scoring is low, OBP is low, then the HR does severe damage, hence, we REALLY REALLY want HR hitters.

So, the entire balance about OBP to SLG is completely dependent on the run environment.  In any case, we should NOT use OBP or SLG.  That's why we have Linear Weights.  Use those.  OBP and SLG will just mess you up.

Friday, May 17, 2013

Scaling WAR, Win Shares, and The Indis

This is a pure math post.  It's not meant to advance anything.  It'll just give you an understanding as to how everyone's methodologies can tie together.?

Read More

Friday, May 10, 2013

Cloning players

In response to a reader at Bill's site, I offer:

I think the reader was asking the following: You have two players of drastically different skillsets, but if surrounded by 8 typical MLB players, will have the same overall effect (i.e., both teams will score an identical number of runs).  But, if you surround each of these players with 8 clones of each, which team will score more runs? 

So, I took two players with the following OBP/SLG profiles: .386/.306 and .293/.473.  Both players have the same 1.8*OBP+SLG values, which is a simple way to get equivalency (assuming they play with 8 typical MLB players).  A team of these high-walk, low-HR players scored 10% more runs than the low-walk, high-HR players. 

I think the reason is obvious: if you have a team of all-HR hitters, who are they going to drive in?  It really kills the value of their HR.?

Tuesday, May 07, 2013

Reader Mail of the Day: Power metric

A reader sent me an email, asking about a power metric, and he touched upon the different ways to measure it, using traditional stats.  Basically:

1. TB divided by H (or equally SLG/BA, since both use the same? denominator, you get the same result).  In this case, we are simply looking at hits as the opportunity factor.  If you have 200 TB in 100 H, it doesn't matter if you did that in 300 AB or 600 AB.  The outs are irrelevant.

2. (TB-H)/AB, or ISO or SLG-BA.  All are the same thing.  In this case, our opportunity factor is at bats.

3. (TB-H)/(AB-SO).  Similarly, it's trying to find the opportunity factor, and in this case, it's contacted plate appearances.  And as #2, we remove the first base.

4. SLG/wOBA.  This one is a bit more odd.  I don't like this one, because it says that the more a hitter walks, the less power he has.

Anyway, I don't like these approaches, because it starts with the answer.  There's no reason being given for what is intended to be measured.  It instead asks: "what does it mean if I combine the metrics in these ways?".  It's not the best way to create a metric.

So, forget all the numbers, and forget all the metrics, forget that SLG and ISO even exist, and simply ask the question as to what you need.  THEN, we can figure out how to create an appropriate metric.

Monday, May 06, 2013

Get to know a stat: wOBA, wRC

Nothing new for the regular readers here, but I like to point out to blogs who take the time to spread the word.  Plus, I like their logo.

In the past, when ?I described wOBA, I usually went with these weights:

0.7 BB, HB

0.9 1B, ROE

1.3 2B, 3B

2.0 HR

I like things to one decimal place, I like to group things, I prefer to keep the coefficients stable, and I prefer to treat the speed portion of the 3B separately from the 2B.

Given that you are talking to people who are aware of SLG (and the 1,2,3,4 weights of the 4 types of hits), and those same people are aware of OBP (and how a walk is as good as a hit), and they realize that both metrics have something that just doesn't feel quite right, though they never thought too much about it, taking the next step to wOBA just seems like a very small step.  I think it should be an easy sell, if they've already decided that SLG and OBP can be improved upon.

***

My favorite part of wOBA is that it really is identical to Linear Weights (per PA) from Pete Palmer, but never (and I mean never ever) have I had the problem that Pete had with negative numbers and 0 being league average.  It just doesn't come up.  I think centering the scale around OBP, rather than around 0=league average, is what resonates the most with the fan.

(Yes, yes, I know some of you want batting average.  Too bad.  Batting average is so overused relative to OBP that I'd never do anything to give it even more visibility.  Batting average is a periphery stat, like BABIP or something like that.  Actually, it's worse than BABIP.  OBP is a paradigm shift from batting average, and the less we use batting average as a reference, the better.  If OBP had never existed, we'd have to invent it.  If batting average never existed, it would be inventing and lie down in the dustbin of history, a cute little metric to be used on occasion.  But one where we'd never consider removing SF from the denominator.)

Saturday, April 13, 2013

WAR v Rosenthal?  No, I wish.  It’s actually WAR + Rosenthal!

Rosenthal highlights Ben Zobrist, who has picked up the "player that no one will remember in ten years as being best for a time" mantle from Chase Utley.  And he peppers it with plenty of good quotes from Forman.

Does someone want to try to come up with that list historically?  Led the league in WAR for a 4-year time period, but never finished in the top 3 in MVP (or top 2 in Cy Young) in any of those seasons.  Does that work?

Thursday, March 21, 2013

Relevance of RBI

When was the last time you learned something about baseball (the game or its players) that you could? only do so because of RBI?  Or in other words, what did RBI offer you that you couldn't get elsewhere?

Related thread.

Saturday, March 16, 2013

Newt and I see eye-to-eye?

?This is basically how I describe the WAR framework, to distinguish it from the various implementations of WAR (that you'll find at Fangraphs, BR.com, etc):

We don't need new principles, but we need lots of new ideas about how to implement those principles in the 21st century

Tuesday, March 12, 2013

Hidden Value in baseball stats

Lee does a good job pointing out that outside of BA, OBP, and SLG, you should also look at DP, reaching on error, and strikeouts.  Which is why you should take a longer look at RE24, which both Fangraphs and Baseball Reference have.  Rather than treating every single non-hit, non-walk the same, it actually tracks what happens to the batter and runners (if any) for the play in question.  In addition, each hit and walk are also treated differently, since a HR with bases empty has less impact than with men on base.?

Friday, March 08, 2013

Correlation is not causation, example #5,673,923

This was posted, with the following tagline:

"During some initial research we were very surprised to find that in 2012 a team walks were not statistically related to runs scored."

First off, that is a positive slope.  If it was unrelated, we'd get a line parallel to the x-axis.  However, the slope looks like it's about 0.15 runs per walk, when it should be double that.

Secondly: if you build a team with plenty of walks, intentionally do that, targetting that, you may do so at the expense of power. 

So, it's possible that what we are seeing is that teams that have lots of walks have fewer HR or extrabase hits.

You are going to score as often on a walk as you will on a single.  If you are showing data that somehow disagrees with that, you are likely seeing some bias that you didn't account for.  Which is what's happening here.

?

PythagenPat Exponent at Various Run Levels

I'm about ten years late in doing this. 

All I did was run the Tango Distribution at various run levels.  The x-axis is runs per game for the two teams (i.e., RS + RA).  And each line is various combinations of RA/RS.  As you can see, in the run environment that is most common in MLB history, around the 8-10 runs per game mark?, the exponent is fairly flat.  Of course, if we try to apply it to pitchers like Pedro in a high run environment, or Koufax in a low run environment, that exponent changes significantly.

Another way to look at it is by the runs per win rate, at those same run levels.  (Note the y-axis is logarithmic scale in this case.)  We see a pretty strong relationship between runs per game and runs per win.  And in the historical level of 9 runs per game, it takes around 10 runs per win.  But we see at the low run environment, the curve is much different.  At 3 runs per game, it's between 4.6 and 4.9 runs per win.