THE BOOK--Playing The Percentages In Baseball

Quoting your article: “And instead of using this delta approach with run values, we use win values.”

What is a “delta approach”?

The difference between the positive event and the out.  I’ll make sure to note this better.

Tango, content-wise, this is excellent.  I learned a lot from reading it.  Gives a great perspective.

I do think it’s pretty dry and technical the way it’s written right now.  I might slow things down a bit and provide more explanation, particularly in the “Now the fun part” section.

When you write something like this…

Everything that a batter and pitcher needs to know about how to approach an at bat is captured in the above chart. Focus on the runners on 2B, 3B, with 1 out. What do we see? The weights of the single and extrabase hits jump substantially, while the weight of the HR jumps a bit.

...cite the exact numbers you’re referring to, to help the reader see what you mean.

I get confused in the last two sections (Game State and Leverage Index).  I mean, I see what you’re doing there, but it’s confusing to read.  First, you add the game state (saying we “need to know it"), then you take it away with Leverage Index.  I think you can set it up better, explaining to the reader what you’re doing and why.

In fact, maybe these two sections should go in Part Two (is that what you already have in mind?) and you should expand the first two parts (wOBA and the Fun Part).  You can end by saying “wouldn’t it be nice if we had a stat that reflected wOBA by situation?  Stay tuned.”

Yes, I was thinking I need to expand further, just about everything.  I wrote it in an hour, so clearly I wasn’t trying to be too careful.  I’m already working on expanding the wOBA by base/out state.

Thanks for the thoughts.  If anyone else has other issues or confusion, please let me know.  And splitting it as a two-parter might be beneficial.

Yes, splitting this into 2 parts would be beneficial. You are presenting a lot of information here. Perhaps you could do the first part on run values and the second part on win values.

It is the single best way to evaluate the performance of a hitter, given the base/out state.

Am I reading your chart correctly and understanding what you are trying to state correctly that you are claiming that hitting a home run with two outs and nobody on base gets almost twice the value of hitting a home run with two outs and the bases loaded?  If so, I think I will stick with BRAA which has the 2 out grand slam weighted at 3.316 and the solo 2 out homer at .891.

Peter - I believe that is the extra value the HR as in that situation, compared to all other possible outcomes.

With 2 out and no one on, there’s very little chance of scoring. A HR puts an immediate run on the board. With 2 outs and the bases loaded, even if the batter doesn’t hit a HR, there are is still a high expectation of scoring a run somehow.

So the relative value of the HR is less. With no one on, you need a HR.

I was confused by the last paragraph, talking about WPA/LI. Everything else was crystal clear.

Brian - I gathered that it was relative value against all other events, but then Tango says what I quoted above, which is what confuses me. I am not sure that a pitcher or hitter would judge the two out solo home run either more “important” or more “valuable” just because it has greater value against the total of all other events. As Tango correctly points out, knowing the relative value is important for determining what strategies the batter and pitcher may be employing during the at bat, but I don’t think that that logically implies that success in implementing your strategy is the best way to “evaluate the performance of a hitter.”

Perhaps the error is looking at success versus failure (HR versus Out) as a ratio instead of absolute runs gained.  The ratio may determine the strategy, but the value of the performance is the value of the runs (in absolute terms) that you either gain or lose for your team by your performance.

Peter, what Brian said.

If we simply want the actual run value, we have Table 50 in The Book, or similarly this one:
http://tangotiger.net/RE9902event.html

And in this chart, with bases loaded, 2 outs, the run value of the single is 1.5 runs.  With bases empty, the run value of the HR is 1.0 runs.

But, if you treat each base/out state as its own universe, then the bases empty 2 outs situation has extremely great impact for the HR.

So, before we decide which of the two charts we want to look at (linear weights, or wOBA), you have to decide what it is that you want.

The wOBA chart is simply the linear weights chart scaled so that the weighted sum of the positive events is exactly equal to 1.

Peter, I don’t disagree with you.  But, remember my quote, which has the condition.  Given the base/out state, it’s the best way to evaluate the performance of the hitter.  So, within that universe of base/out state, that’s how you do it.  What it does is scale each base/out state so that they are all with the weighted sum equal to 1 for the safe events.

Otherwise, you can have someone with +12 runs in bases empty 2 outs per 100 PA, and you can have someone +15 runs with bases loaded 0 outs per 100 PA, and it’s hard to see how to evaluate this hitter.  On a wOBA scale, in the former situation, he may have been unstoppable, while in the latter, he was merely great.

Two typos in the article, Tango.

It might be worth discussing or at least mentioning the possibility that batters have absolutely no control over their approach.  It is possible.  We are pretty sure that they do and we are positive that each batter’s profile has an effect on his value, given the game state, but we really don’t know how much of the difference between a batter’s regular lwts or wOBA and your WPA/LI is due to luck, and how much is due to skill. IOW, how much is repeatable or predictive.  I hesitate to use a non-context neutral stat over a context-neutral one, without at least having some idea of how much luck/skill there is the former.  If it is small, then what’s the point?  At least from a going forward perspective.  If its only value is from
a retrospective perspective, then we have the same debate about what constitutes value in the past.

I don’t want people to tell me that player A is “better” than player B because, “Look, even though his regular wOBA is much worse than player B’s, his WPA/LI wOBA is much better!” unless I know that there is a large amount of skill there, any more than I want someone to tell me who the best pitchers are based on their BABIP over some time period.

Anyway, rather than discuss the merits here, I think it is worth mentioning and perhaps even discussing, in the article. You say, “the single most accurate metric” or something like that.  Considering the skill/luck issue, that is highly debatable from a semantics perspective.

To me, the value in these numbers and this construct is in teaching hitters and pitchers what kind of approach to use ("Look, what the value of each event is in this situation!")

Just telling me each player’s WPA/LI is not particularly interesting to me, at least until you prove to me that there is a large spread of talent in this metric independent of a context-neutral one like regular lwts or wOBA.

I’m going to have to repeat this:

It is the single best way to evaluate the performance of a hitter, given the base/out state.

Note the qualifier.

If you have runners on 2b and 3b and 1 out, it is unreasonable to use a context-neutral linear weights formula over a context-specific one. 

Now, for some of those states, the change in relative values might be too nuanced for the batter/pitcher to appreciate, and therefore, won’t bother to change their approach.  That’s why a week or two ago I suggested collapsing the 24 base/out states into 8 base/out classes.

So, while I may agree with you that we shouldn’t use 24 different lwts equations, I have to disagree that one will do.  I don’t think it’s up to me to prove that, but you to do that.

***

No, I don’t think it’s possible that a batter has no control.  There’s no question that the pitcher must approach that situation differently.  Given that the pitching is pitching differently, the batter is going to respond in kind.

We know pitchers throw differently based on count.  I don’t think it’s a stretch at all to say they throw differently based on base/out or game state.  And, I don’t think it’s a stretch at all to say the batter responds to that.

I think at the very least, having 8 wOBA equations will respond to that.

That said, definitely we can have a fantastic dialogue and research on this topic.  This article can certainly serve as the initial basis.  Indeed, this is what baseball is all about.

If you have runners on 2b and 3b and 1 out, it is unreasonable to use a context-neutral linear weights formula over a context-specific one.

No it is not unreasonable, but it may “mean” absolutely nothing (if the ability of the batter to change his approach over and above his overall talent is little or none, as compared to the inevitable noise in the metric) in the grand scheme of things, whereas using one static set of lwts value may mean a lot - namely a sample of the player’s overall offensive talent!

I think you need to at least mention that in the article, but that is up to you.  It is a good one (the article) either way.

When you mention that something is the “single best...” without mentioning the possibility that much of the distinction among players may be noise, you are going to be misleading people whether you mean to or not.

The burden is on YOU to show me how much the distinction among players is noise is how much is not.  Otherwise I can spout numbers all day long to tell you who are best day or night players, home and away, versus certain teams, etc., with the clear implication that these “splits” have predictive value.

You are essentially giving us “splits.” With splits, you MUST tell us how much of the variation among players is noise and how much is skill, otherwise you are going to be raising a generation of fans who believe that there are real live “lefty mashers” and pitchers like Dempster who are “road team killers.” Etc.

It is of course a fair point that since we are looking at samples of data that we need to regress that to get to the true talent.

MGL’s argument basically boils down to this: that he can tell who is a better hitter in the base/out state of 2B, 3B 1 out by simply looking at his overall batting line, than by looking solely at his performance line in that particular base/out state.  And the reason is that in one case he has 600 PA, while in the other he has 30 PA. 

So, he’s absolutely right that my statement needs to be qualified.  At some point, the size of the sample for a particular split line will exceed the reliability of the overall performance (non-split). 

For example, if you tell me what the park factor is for a LHH in SF home park (3COM or AT&T or whatever corporate name it has now), and I have Barry Bonds actual split line after 500 games in that park, I’ll tell you to go take a hike.  There is simply no way in the world I will pay attention to that split line other than to put in the mildest of regression.

So, it would be ridiculous for me to assert that after 30 PA, I can tell you how good a hitter is in a particular base/out state.  And clearly his overall performance across all 24 base/out state would be weighted more than the individual split line.  But, at some point it will matter.  What that point is, I don’t know.  I can guess that after say 400 or 500 PA in a particular base/out state, that that will trump the overall batting line over the same time period (i.e., both equally weighted).  It’s possible that say for a specialized base/out state like man on 2b, 3b, 1 out, that it’s less and that in a ho-hum base/out state that it might be alot more.

So, if this is what MGL is talking about, I will accept that his position is justified.

Ok, I updated the article.

It’s got alot more technical stuff and charts, so maybe that might clear things up here.  If you guys are good with this, I’ll expand it even more in terms of explanatory text, and send it to THT.

I will also grant you that when adding it all up, which presumably is what you are trying to do, you are going to get a more meaningful number than just looking at 30 PA in one particular base/out state.  In fact, that is what I meant.  Not that a particular base out state is not going to tell us anything useful as compared to his overall performance, but that even that adding everything up (across all base/outs states) might not tell us more than an overall lwts or wOBA using the same coeeficients for every base/out state.  I don’t know.  I am just saying that you need to see how much, if any, talent there is in an entire season of this kind of metric.

For example, let’s say that we used your lwts by base/out state metric and player A was +10 runs in 600 PA and player B was +5.  But using regular lwts (static values) player A was +5 and player B was +10.

Who is going to give me better value next year across all base/out states?  I don’t know the answer.  We need to know how much to regress one towards the other.  If players have only a little control over their approach (I really should say, if the spread in “control talent” is small - players could have a lot of control, but the talent among players with respect to that control could be the same), then most of the difference between the +10 and +5 for each player is noise, and player A is really close to a +5 (his regular lwts value) and player B is really closer to a +10.

Bottom line, without knowing the ratio of signal to noise in ANY kind of “splits” I don’t really want to hear about those splits.

Kind of like inherited runners’ stats for relief pitchers. A lot of people use them, even the sabermetric sites like BP and THT.  They bore me to death, as I know that it is a meaningless stat as far as any predictive value goes.  Yet people use them to tell you, at least partially, how “good” certain relievers ARE.

Little different application, but likely the same math.

I read an article in the last week or two that looked at individual batter/pitcher matchups, running statistical tests to show when the results are significant.

With those test results you may be able to determine an amount of regression where rhb-a hits x off all rhp, and rhp-b gives up y against all rhb, and when a has actually faced b the result was z. That way you could work in the batter’s overall ability, the pitcher’s overall ability (both large samples) with their actual head to head, to give the manager something more meaningful to put on his index cards.

Tango, this is very dry, technical and hard to read.  And long.  As you add explanatory text, I’d suggest adding an intro paragraph or two that lays out what your intention is and why.

Plus, you have lots of sentences that say things like “Now, what interests us is the run value above the out.” But it’s not clear to the reader why that interests “us.” Presumably, that’s the type of explanatory text you plan to include, but I thought I’d mention it.

A second point is that I don’t think all the tables help.  It is hard to read text with a lot of tables inserted, particularly on a computer screen.  And many of these tables don’t add anything to the reader’s comprehension.  They’re just there as reference.

I also think Table 3 is confusing because, before this article, Leverage Index has been based on an entire situation, not just base/outs.  I would label it a “special version” of Leverage Index.  This becomes a key issue when you refer back to Leverage Index at the end.  Which Leverage Index?

http://www.fangraphs.com/blogs/index.php/best-pitching-performances-5-1/

My reply:

Paul captures it perfectly here.  WPA captures the performance relative to the timing.  Every time the pitcher can get out of an inning not allowing a run, then he keeps gaining WPA, starting with 0.05 wins in the first inning, and going up in the ninth inning.  The amount gained is directly proportional to the Leverage Index at the start of his inning.  Thus, he gains the most when the score is close.  I would not say necessarily that it’s when it’s a 0-0 game.  It might be a 1-0 game.  I’m not sure.

WPA/LI (aka Situational Wins) is perfect for hitters, since it values each PA as “1”, which is really what we want to do, to make a fair comparison.

However, what we *really* want is something in-between WPA/LI and WPA for a pitcher.  A pitcher doesn’t just come into a PA like a batter, but he creates the environment.  And so, each PA must be worth more the more runners are on base.  We really need to multiply the WPA/LI by the LI for the base/out state.  A bases empty situation has an LI of 0.7, while one with the bases loaded is say 3.0 times.  That’s what you really want to do: WPA/LI * boLI.

For a season, a pitcher’s boLI will come in pretty close to 1.  Even if it doesn’t, the boLI of all the top pitchers will be close to each other so that they will all be similarly biased.

For a single game however, it might make a difference.

Not sure if this is the best old thread to bring back, feel free to move it if you have a better one.  I just had a few things to say about situational wins and would prefer to post here than on a recent BTF thread where the quality of debate is in the order of “WPA is teh stupid”.

Here are a few ways to value wins, what you are measuring, and the implications of each:

WPA - Looks at the win impact of an event, at that point in time only.  Player hits a grand slam to win a game 4-3, he will receive more credit for doing so if he does it in the bottom 9th than in the bottom of the first.

Approach #2, Take timing out of WPA. Those two grand slams are equal, but if you are looking backwards, then a guy who hits 3 homers in a game that his team loses would have no value at all.

Context neutral - treat all events the same regardless of when they occur or the team outcome.  You treat the grand slam with 2 out in the bottom of the 9th to win 4-3 as equal to the grand slam to cap off a 21-2 blowout.

I don’t see how anyone can say one approach is better than the other, it depends on what you are trying to measure.  Did I miss a system of win valuation that doesn’t fit into those 3 categories?

I should add that I would count WPA/LI as a context neutral approach.

The WPA/LI is a “nuanced” context-neutral approach.

The most extreme position would be a bases-loaded tie game in the bottom of the 9th, two outs.  In that case, a walk is equal to a HR.  The “WPA/LI” in that case is +.054 wins if a run scores, and -.027 wins if the inning ends.

In a purely context-neutral approach, the HR gets +.130 wins and the BB gets +.030 wins.  Since the batter, pitcher, fans, and everyone realizes that the HR=BB in this case, and the batter/pitcher are changing their approach entirely with the understanding that a HR=BB, I see no reason at all to say that a HR should count as +.130 wins here.  The batter has cut his swing down.  The pitcher is making sure not to walk the batter.

They key point about WPA/LI is that every positive event is centered around +.054 wins.  A walk’s win value will be anywhere from +.02 or so all the way up to a maximum of +.054.  A HR can go from +.054 to maybe +.20 wins.

Setting those kinds of extreme cases aside, the win value of the walk will almost always be .03 wins +/- .01 wins, and the HR will almost always be +.13 wins +/- .02.

WPA/LI really keeps everyone happy.  The problem is that it’s a bear to explain.