THE BOOK--Playing The Percentages In Baseball

It depends on what you are trying to do. If you are comparing players in the same league I think a quick, standard formula might be fine.

But if you are comparing players across time and space then I think we’d want a formula that takes into account the different value of different events given the era in order to make a fair comparison.

That being said, if I can adjust wOBA for the league environment I am looking at I do, even if it is a same-year comparison.

I guess I should also ask: how do you want your FIP?

That is, do you want wOBA to follow the same generic-type FIP formula, or do you, for some reason, want one with more accurate coefficients than the other?

If finer coefficients create greater separation between pitchers, then finer is better. If the difference is neglagible, however, than the easier static FIP might be just fine.

This is a crucial question to me.  Said differently, do we want to have a stat in which we either:

- Trust a website to calculate coefficients correctly, and/or…
- Have the chops to calculate them ourselves

Or…

- Have a stats that’s easy to understand and implement, understanding that it’s not “perfect” for cross-era calculations.

Personally, I vote for the latter.  I don’t like trusting another website to calculate the proper coefficients all the time (witness how BRef and FG can’t even agree on WPA, even though they’re both using Tango’s tables).  And I don’t have the chops to do it myself.

If we still had stats at THT, we’d be using FIP with the adjusted coefficient, and we’d probably add wOBA, with regular weights and a coefficient.

I think it comes down a research mindset or a communication/education mindset.  I say: let researchers do what they want, but make publicly available stats easy to understand and interpret.

I tend to agree with studes here.

I look at something like SLG.  The weights are ALWAYS 1,2,3,4, even though we know that those weights can’t possibly hold in every run environment.  And the weights in OBP are always 1,1,1,1,1 even though we know that in some run environments, those weights make more or less sense.  And we know that just doing OBP+SLG doesn’t always make sense, or 1.7*OBP+SLG doesn’t always make sense.

BUT, it’s nice to see that SLG is always calculated the same way, and OBP is always calculated the same way and 1.7*OBP+SLG is always calculated the same way.

To that end, I’d think that having 0.7, 0.9, 1.3, 2.0 as the standard weights in wOBA would be a good thing.

I think Pete Palmer recognized this with Linear Weights, even though he knew that each era should have different weights.  And Clay Davenport had EqA with standard weights, and he worried about the conversion on an after-the-fact basis.

I think that’s why FIP is as popular as it is.  13, 3, -2, all the time, no matter what.  If we want to be more analytical about it, we can.  We can worry about that later, when it matters.

Tango/5

Obviously, you created wOBA, so it’s pretty hubristic to say what I’m about to say, but I’m going to say it anyway.

The analogy of wOBA to OBP and SLG is wrong.  It’s wrong because the weights of wOBA are supposed to mean something specific.  They are derived from linear weights, which are the differential impact on run scoring of each event.  That’s a function of run environment.  Differential wOBA is supposed to convert to marginal runs, whereas, the weights of OBP and SLG measure bases.  You can’t do anything with OBP or SLG other than measure bases.  So, it’s ok, a base is a base.  But wOBA is supposed to measure runs, and if you don’t adjust the coefficients properly, it doesn’t.  This is why wOBA should have the right coefficients.

Now, wOBA isn’t exact, because marginal run impact isn’t precisely linear in each event.  So, in that sense, it isn’t exact.  But getting the right linearization of marginal run impact seems to be more important to me than making it easy to calculate by hand.  By computer, I don’t think one is any more difficult than the other.

The number one thing I want is transparency. I don’t mind if you use custom linear weights or neatly rounded approximate ones, but I would like to know where you (editorial “you") get them. This kind of goes with Studes #4: I am happy to trust fangraphs, especially if they publish their weights.

If we’re going to go the standardized route, I would prefer a stat that expresesit in terms of a standardized baseline. Maybe something like wOBA divided league wOBA times .330, or times 1. This means I always know how a player compares with his peers.

I realize it doesn’t make much difference, but treating doubles and triples the same just bugs me.

I know you disagree, but I still prefer David Smyth’s basic wOBA—I think it’s actually easier to remember just one constant (0.36) along with the basic weights of 2 for BBs, 1.5 for hits (and ROEs) and 1 for total bases.

(Plus, if you don’t care about the scale, you can ignore the 0.36, which makes it even simpler.)

Wow, somebody actually paid attention to the post where I mentioned that quick wOBA! Personally, I use that formula all the time and it never steers me wrong, and I never use the .36 multiplier because I don’t care if it’s on the OBA scale…

Larry/6: if you support that, then doesn’t that mean that you can’t support the current version of FIP?

phil/7: that’s what wRC+ is for.

The more I think about this, the more I want linears weights tailored to the run scoring environment.

For quick approximations, we have OPS and a quick mental math of 2.OBP+SLG. wOBA doesn’t get much beyond the latter, so why not make wOBA go as much beyond the latter as possible by providing context relative information?

Why does wOBA have to be tailored to the era if wRC+ already makes that adjustment for us? Why make them say the exact same thing?

The advantage of making them say the exact same thing is for people who prefer one scale to the other.

***

I’m more interested if people are going to be consistent with their reasons to keep wOBA as-is (i.e., context-specific) as they’d like FIP to have their coefficients be changed every year.

Tango/10:

Yes.  FIP should behave the same way, in my opinion.

That said, I’m a little less clear on what FIP is supposed to mean exactly, because it explicitly is making assumptions about what happens on balls that aren’t HR/BB/K.  If your question implies that FIP is a tallying up of the marginal run impact of certain events, then I’ve actually learned something from this. I’ve seen you resisting the idea that FIP is a true-talent estimator recently and was kind of confused.  I believe I understand now. 

If that understanding is true, then my answer changes to “Absolutely.”

I had a thread a week or two ago ("Deconstructing FIP") that shows that the FIP coefficients are dependent on the run environment of that pitcher. 

If I do that though, then the whole “love” for FIP goes away.

Similarly, perhaps wOBA should be more “love” and less “pure logic”, so that it gets the same kind of respect that FIP gets.

"Similarly, perhaps wOBA should be more “love” and less “pure logic”, so that it gets the same kind of respect that FIP gets.”

This is why I’m still a GPA fan.  The only GPA fan, yes, but a loyal one.

If we’re mainly talking about why FIP has gained more acceptance than wOBA, I don’t believe it has anything to do with complexity.

wOBA has a not so distant cousin: OPS.  OPS doesn’t get the weights right, but it’s not the most horrible stat in the world and does a half decent job at what it’s attempting to do.

FIP has no relatives in the mainstream and happens to be one of the simplest along the DIPS line of thought.

We can use SBNation search as a proxy for acceptance:

On SBNation there have been 32,226 comments that include the word FIP and it has been mentioned in 7,970 articles.

wOBA/wRC+ has 24,291 comment mentions and 4,759 article mentions.  For reference, OPS/OPS+ has 140,981 comments and 23,492 articles.

UZR just for kicks: 27,524 / 3,396

And here’s an absurdly simple yet quite effective stat—kwERA: 17 / 3

To me, these are pretty decent numbers for wOBA.

In my opinion, you can really calculate wOBA however you like.  If you want to calculate it on your own using simplified weights for back of the envelope calculation, I don’t see any problems with that.  The results are going to be pretty similar.  If you want to use yearly weights and work with the SQL that tangotiger originally provided, you’ll get the exact same results that are on FanGraphs.  If you want to make things even more complicated and try to get even more exact weights, I doubt anyone will take issue with it. 

Either way, no one is calculating OPS or wOBA in their head.  You at least need pen and paper (at least I do).

There is a point where simplicity and diminishing returns meet in my mind and more importantly in the publics mind.  It’s a gray area and I don’t have any personal rules, but for me it’s usually when something is just too annoying to code that I wish for something simpler.

If I could create a stat tomorrow that was 100% accurate and represented perfectly what a batter did at the plate with zero flaws, but it took me 10 million lines of code and an entire data center of processing power to calculate (but was still easily accessible just like wOBA) do you think it would catch on?  I say it depends.  Is it 1% better than OPS or 50% better or even 90% better?

Complexity is not wOBA’s problem.  It’s OPS (or maybe even more importantly it’s SLG and OBP).  If anyone is interested in furthering wOBA’s mainstream acceptance, it’s going to take years of work to supplant it.  FIP is really up against nothing except newer and less well known ERA estimators.  It was in the right place at the right time and does a good enough job for most people’s purposes.

Yet wOBA has shot through a certain segment of the online sabermetric community.  Why is that?

I would definitely go for adjusting it to the RSE, except that it’s even more than RSE: scoring 4.2 runs/game based on a higher HR profile or a higher BB profile are two pretty different things. But I’d also want to adjust this on a team-by-team basis, and, and, and, and… well, I might as well just come up with something else for all that (I’m a bit of a precisionist), so probably the simple is better.

There’s a certain segment of the baseball community that recognizes and cares about the flaws that are in OPS and thinks wOBA is a worthwhile and significant improvement over OPS.  If that segment grows large enough (which it might), then wOBA could gain more mainstream acceptance.

There’s a tremendous amount of inertia when it comes to new things.  Why go with something new when you already know the old thing and the new thing is only marginally better and has even a slight learning curve.

I just think that using complexity as a reason for FIP’s success and wOBA’s slightly less success is operating under a false premise.  I think having a simple way to present wOBA or even an “official” formula could be useful, but wOBA has had that from the very beginning:

(0.72xNIBB + 0.75xHBP + 0.90x1B + 0.92xRBOE + 1.24x2B + 1.56x3B + 1.95xHR) / PA

It is a little bit true that using a single constant for each year gives you many less numbers to remember on a yearly basis, but once you start implementing changing constants, you’re no longer doing back of the envelope calculations and it doesn’t matter anymore.  You then need a spreadsheet or a database and the extra work is pretty minimal.

I’ll disagree with you, David.  You raise good points about there being alternatives to wOBA, but I don’t think FIP would have reached anything near its current level of acceptance if Tango had first introduced it with changing coefficients, or coefficients that weren’t round numbers.

It’s not a matter of whether or not a spreadsheet is required; heck, I can’t do the math of any metric in my head.  It’s how simple and understandable the metric is.  FIP is simple in construct and has just three coefficients.  Nice, round numbers, too.

wOBA doesn’t have that, even in its simple form.  Heck, maybe David’s formula is the way to go.

I agree with both studes and appelman.

Appelman makes a good point about alternatives.  But I concur with studes about the non-changing simple numbers.

I also agree with the third David and Craig that: 2*BB + 1.5*H + TB per PA is a nice simple round thing to use.  It’s extremely close to wOBA.

I should note that this is the core of EqA’s numerator:

H + TB + 1.5*BB

As you can see, Clay sort-of got the coefficients correct.  Bump up the BB and H, and hold TB, and we’re in the same boat.

So, that’s what Clay did, is that he keeps the core intact, and then worries about the adjustment on a more global level (like current version of FIP does).

The fact is that each component should change by run environment, so, it’s a matter of how precise you need things done.

One other thing about this issue.  FIP has a big advantage in that it’s being compared to ERA, which is a (relatively) universal way of comparing pitchers.  ERA is very useful, and FIP is a good alternative way of looking at a pitcher, using the same scale.

Batting is different.  There is no one scale to build upon.  There was BA and RBI, but we’ve managed to move some folks beyond BA and onto OBP.  I think that RC is (was) a little bit successful because it could be compared to R or RBI.

But I doubt that batting will ever have a success story like FIP.  It’s not just that there are good alternatives, it’s that the basics don’t line up.

studes --

You can scale the simple wOBA to batting average if you use 0.29 instead of 0.36.

I initially was a big proponent of scaling to batting average on familiarity grounds (and I note it was one feature of GPA).  But scaling to OBP makes more theoretical sense—it really is a weighted on-base average. 

Also, I think OBP is gaining greater acceptance.  On some telecasts, it’s presented along with batting average.  And when my sons argue over the relative merits of two players, they’re just as apt to bring up OBP as batting average. 

(As an aside, I was quite proud the other day when my younger son responded to his brother—during a heated argument about Casper Wells— with the retort, “That’s a small sample size!")

The denominator of OBP is the same as wOBA (IBB notwithstanding).

The denominator of BA is an abomination.  I just don’t like the idea of scaling something that is per PA into batting average’s per AB.  It has no appeal to me whatsoever (and I’ll never do it).

I didn’t mean to bring up scale as a big issue.  I don’t think it is.  My point is that FIP has received greater acceptance partly because it matches well against a singularly referenced pitching stat.  There is no singularly referenced batting stat.

I can understand people liking the simplicity of FIP with its integer coefficients.  But for some uses we would want as much accuracy as possible.  Does Fangraphs use basic FIP, or a more complicated version with variable coefficients in figuring fWAR?

It uses basic FIP.  Now that you bring it up, that should be corrected.  FIP is a shortcut for a more elaborate BaseRuns model with league average BABIP.  So, great point there.

Now that my quick wOBA has been mentioned a few times here, I might as well explain why I like it. First, Tango mentioned that it’s “extremely close to wOBA”. Second, although it’s written for convenience as (1.5*H + 2*BB +TB)/PA, it’s actually better conceived of as (2.5*H + 2*BB + [TB-H] )/PA. That is, it separates basic batter offense into hits, walks, and extra bases. And that, for me, pretty much nails the essential and different skills of various batters. Pujols gives you mucho hits and extra bases, and decent (non-intentional) walks. Ichiro gives you hits, and that’s it. Dunn (pre-2011) gives you extra bases and walks, but few hits. And so on.

Whether that tickles your fancy as much as it does mine is for you to decide. I am quite comfortable with that framework, even if it treats all extra bases the same, which seems to be the main difference relative to full wOBA. The older I get, the more I understand the complexity of the world, but also the more I appreciate the benefits of simplicity whenever possible.

dave/30:

Not to pick on you, but this is exactly why I think it is important to get it right.  With these coefficients, you could start extrapolating on the relative values of these different skills, comparing different players with different mixes, and so forth.  But, all you are doing with that is propogating the erroneous coefficients you started with.  I think that’s going to lead to bad conclusions.  Maybe the differences aren’t enough to matter, but I think when you start doing things with them, you won’t know when the differences begin to matter or not.

Larry/30, just where are my coefficients “erroneuos”?

Well, all “extra bases” are not created equal, so it is necessarily an approximation of wOBA and not an exact rendering of the value of the actual batting skill a player brings to the table.  That’s what I mean by erroneous.  Perhpas I should say “approximate” instead, or maybe “biased.”

David’s coefficients, when multiplied by 0.36, gives us this:

0.72 BB
0.90 1B
1.26 2B
1.62 3B
1.98 HR

That’s as good a set of weights as you’ll find.  The 3B is a tad high, but if you’re going to be high, might as well do it on the low frequency one.  The error on that is .003 in wOBA points for an extra triples hitter.  We can live with that.

I mean extreme triples hitter.

By “skill”, here I mean only the actual distribution of events, which is what wOBA measures, not the expected future distribution, which is something else entirely.

In any case, what I’m getting at is that the derivatives of wOBA should be the appropriate linear weights.  Otherwise any analysis using wOBA is going to get inaccurate answers.  Maybe wOBA shouldn’t be used for that purpose, that’s what RC or RC+ are for.  But, I was under the impression the whole thing was supposed to hang together.

Tango/34

Fair enough.  But wasn’t the whole point of this discussion changing run environment?  Doesn’t this sort of thing necessarily build resistance to doing the correct thing with respect to run environment?  I suggest yes.

If we go here:

http://www.tangotiger.net/bdb/lwts_woba_for_bdb.txt

We see that David’s weights pretty much nail 1994-1997.

For a set of weights that are constant, David’s version fits the bill as well as anything you can come up with.

Well, this is the question.  FIP holds the coefficient constants.  If that’s what people like, then David’s version of wOBA fits the bill.

If we want wOBA to vary the coefficients, then we want FIP to vary as well.

All depends what people want.

Question.  Is there a “preferred” way to set weights these days?  We’ve gone through several different approaches over the years (regression, Markov chains, Base Runs, etc. etc.) that I’m out of touch with how folks prefer to set their weights.

Are most of these sites/people adjusting for parks?

I like what Tango had on the deconstructing thread, wOBAfip.

Try this (0*K)+(.3*BIP)+(.7*BBHP)+(2.0*HR)/PA

I like the simplicity, the predictive element, and the ability to keep it the same for hitters and pitchers.

I have been using the above formula to analyze players recently. Believe me it stands up pretty well.

Right, that is pretty lovely.  Do we really need that 2.0 to be 1.9 or 2.2 depending on the run environment?  Do we really need that 0.3 to be 0.28 or 0.31 depending on the run environment?  And so on.

So, I think there’s a very nice reason to have non-changing coefficients.  I think it leads to broader acceptance.

But when you need to calculate things for analytically purposes, then you should go to the trouble of getting it right.

The FIP thing in fWAR is a good example.

Can’t we just have two different stats? Traditional wOBA and the new one? Same for FIP?