THE BOOK--Playing The Percentages In Baseball

Btw, this is why a stat like Win Shares does not give extra credit to a winning team.  Since Leverage is fairly evenly distributed across all teams, and since wins equals performance times leverage, and since leverage is around 1.0 for all teams, wins equals performance.  So, forcing the totals of your players to equal team wins won’t be unfairly biased based on the performance of the team as a whole.

Thanks, Tango.  I’ve mentioned the second point on my site.  A little later this week, I’ll post an article on WPA and Win Shares so far this year.

I’ll take this opportunity to say that I can’t stand WPA, as now everyone is assuming that the new sabermetric uber-stat takes into consideration hitting (or pitching) in the clutch.

I’ve said this a million times and I’ll say it again.

One, since every saber researcher and their mother have shown that an ability to perform better or worse than the average player in the clutch is at best a small skill that very few players possess, and hence has almost no predictive value, waazz the point of WPA??

Two, “the clutch” is only “the clutch” if the player perceives it as such.  Players as a general rule know when games are on the line, but there are many situations that WPA sees as “clutch,” that a player would not, and vice versa.  For example, everyone must think that a 2 or 3 run lead in the 9th is a “clutch” situation, or why would managers feel a need to bring in their best pitchers?

Three, if you want to come up with a stat to reward performance regardless of whether it was predicated on “luck” or skill or some combination thereof, it makes no sense to reward performance which leads to a loss.  WPA of course, while it rewards large positive changes in WP and give demerits for large negative changes in WP, is silent about whether the game was actually won or lost by the player’s team or even whether a run was even scored or not.

If we are going to reward “lucky” behavior that leads to a win, fine, I have no problem with that.  If we are going to reward true (context-neutral) skill (like lwts) regardless of the outcome, I have no problem with that either.

To reward luck-based performance that leads to a loss is ridiculous in my opinion.

Anyway, I hate WPA and find it utterly useless.  Period.

I’m definitely not with mgl here. 

Who’s to say that David Wright’s context-neutral performance this year is not luck?  In fact, we would have to regress his performance more than someone else who has performed equally as well.  Papelbon, Liriano, etc.  Or Brady Anderson’s 50 HR season (assuming no steroids).

If you are going to do things retrospectively, then I give 100% credit to the impact of the performance to the players involved.  LWTS gives a single +0.47 runs, or +.043 wins, whether that single actually ends up scoring a run or not, or leading to a win or not.  It simply gives it the theoretical change in RE or WE.

But, since players do in fact change their approach in hitting or pitching with men on base or not, I don’t see why we wouldn’t at the very least give out LWTS by the 24 base-out states (Table 50 in The Book). 

Taking the next logical step to WE does imply that players change their approach based on the leverage of the situation.  Whether true or not, I still need to assign +.50 wins for every game to the winning team and -.50 wins for every game to the losing team.  WPA is the way to do it.

Anyway, I love WPA, and find it fascinating.  Exclamation mark!

You may like it, but of what use is it, and I don’t mean that rhetorically.  I agree that lwts means nothing in and of itself either.  However, once you apply the proper regression, it is EXACTLY useful in projecting future performance and value in any context.  It is silly to say, “Who is to say that a player’s lwts is not luck?” as if the same can be said about WPA.  They are apples and oranges as far as luck and skill are concerned.  One is exactly a sample of a player’s skill and the other is not.

I’ve also said before, if you want to use WPA as some kind of an MVP gauge or something like that, be my guest.  However, there are a thousand other metrics that are equally legitimate for evaluating MVP-type performance, depending on what you want to reward.  As I have said before, a player can have the highest WPA in baseball and his team can be 0-162. If you want to give that player the MVP award, I think you would be laughed out of the room.  I cannot think of any other use for WPA (other than it is fun and interesting).  Please enlighten me.

This is about the 3 “temporal” perspectives--real time, future, and past. I wouldn’t say that any of these have “intrinsic” impotance--they have value only if the observer deems them interesting.

And despite mgl’s objections, people do find real-time WPA interesting. People are also interested in the future, of course. As for the past perspective, several years ago I developed an outline for doing exactly what mgl described--starting with actual wins, and giving credit for contribution to actual runs in these wins, etc. It must not have been very interesting because nobody with the requisite PBP data picked up the ball, and it died after a single short thread.

Why do people like the WPA perspective? MGL might say it’s because they haven’t appreciated the significance of the flaws he points out. It’s partly that, but it also may be because the sense of excitement and anticipation associated with real-time observation is being partly transferred to the numbers. When you look at those WPA graphs of a game, you feel like you are reading a story…

For me, WPA is interesting in its analytical structure, and it does give *some* additional info for MVP discussions. But I don’t find it all that compelling.

"This is about the 3 “temporal” perspectives--real time, future, and past.”

That’s a useful distinction.  But the problem with WPA in the MVP/value context is that we are talking about the past, yet using a real time metric that can’t incorporate all the knowledge we now have.  One missing element, as MGL notes, is who actually won the game.  But a lot more than that is missing: a leadoff triple essentially counts as a run scored, even if the batter never scores.  More broadly, WPA misses what the FOLLOWING hitters did.  If a hitter performs with men on base, he gets “extra credit.” But if he gets on base before a hit (and so scores), there is no extra credit in WPA, though getting on base at the “right” time is clearly more valuable than being stranded.  (I realize that it sounds odd to think of this as clutch performance, since a player can’t possibly know what subsequent hitters will do.  But since we know that few hitters can consistently outperform in high leverage PAs, there’s no logical difference in the two situations). 

Evaluating a player’s performance in the context of what was knowable at that precise moment—which is what WPA does—is only interesting if you believe that players differ in their ability to deliver in the clutch.  If you don’t believe that is true, then logically WPA cannot be an important metric for establishing value—it’s probably not as good as R+RBI (and I think we all agree it’s not at all helpful for understanding true talent).

WPA is the only metric that can properly scale hitters, starting pitchers, and relievers.  Now, you can try to get around this by applying a player’s context neutral performance and multiplying it by the LI, so that you get a quasi-WPA.

In this case, you are giving credit (or debit if he fails) to the player for the LI.  That may sound strange, but if the Redsox intentionally keep Papelbon in the bullpen, rather than logically starting him, then that must mean that his true worth is much more than the 80 IP he’ll end with.

LI Tangent

As well, Ozzie Smith ends up with a high +/-, in part, because the SS is involved in far more plays than a 1B.  Ozzie didn’t will those balls to SS, but his talent level allows his manager to leverage his skills.  In essence, the LI for a SS is 1.25, and for a 1B, it’s 0.75 (or some such).  I’d have to give Ozzie credit for being in that position.

Same can apply for hitters, where a leadoff hitter gets 10% more PA than the #5 hitter, and this may have nothing to do with the player, that being the choice of the manager.

Back to the argument

I’m with David, as to the fascination with WPA.  It’s real-time, and numerically expresses our inner feelings as we watch the game.  WPA provides the most interest to a gambler, as he wants to know the chances of some team winning. 

Think of WPA just like with a gambler, or a stock picker.  The monkey can get lucky, and pick the winner more often than not, and get a million bucks for his trouble, while I may be down just as much.  Whether the monkey or I exhibited some skill, I don’t know.  What is real is that the monkey is partying with the hot chicks, and I’m not.

Derek Jeter is partying with the hot chicks, and ARod is not.  That’s what WPA tells you.  That’s real, and it tells you exactly why the Yanks won or lost.  That Jeter performed well in occasions where it counted most, and ARod did not.  Maybe Jeter is really a monkey, and ARod is a finely-tuned baseball machine.  That’s irrelevant to the fans.  The fans want to see the timely production, and they will lavish their love to whoever happens to do so.  Right or wrong.

WPA paints that picture.

"That’s what WPA tells you.  That’s real, and it tells you exactly why the Yanks won or lost.”

But it’s NOT real.  If Derek gets a single and moves a man on 1B over to 3B, he gets a lot of WPA credit even if the next hitter hits into an inning-ending DP.  What’s “real” about Derek’s WPA credit?  RBIs and Rs are real (though flawed in ways we all understand), but WPA is not.  WPA is not context-neutral talent, yet not real value either—and that’s why some of us are less excited about it.

* * *

Tango:  Is WPA usually calculated using empirical WE, or mathematically-generated WE assuming two average teams?  I would think it makes a difference: if the former, a closer is being measured (mostly) against other closers; if the latter, against the “average pitcher.” I’d be interested in hearing your thoughts on which method is better, and why.

I love WE, and find it to be an invaluable tool to evaluate in-game strategy and the like.  But once the game is over, why look at WPA?  It certainly doesn’t measure the player’s ability.  It certainly doesn’t measure the player’s actual value (because, as MGL notes, it rewards positive performance even if that performance doesn’t lead to a win).  What it measures is “expected value,” which loses its relevance as soon as the game is over and we know the outcome.

In other words, the real-time perspective is only useful in real-time.  If I look at a list of individual WPA leaders for the 2006 season, it doesn’t tell me anything particularly helpful.  I don’t know how much of each player’s WPA total is “clutch"-related, nor do I know how responsible the player is for his team’s overall success (or lack thereof).

“As for the past perspective, several years ago I developed an outline for doing exactly what mgl described--starting with actual wins, and giving credit for contribution to actual runs in these wins, etc.”

David, I remember that thread very well - count me as one of the few people that found it interesting.

The trouble with the past perspective is that it’s really tough to come up with a theoretically sound framework.  The crux of the problem is that, in an absolute sense, hitters create wins and pitchers create losses (I’m pretty sure that your system was built around that idea).  But within that framework, how can we differentiate between a pitcher who throws one scoreless inning and a pitcher who throws 200 scoreless innings?  I can’t see a way to do it unless we give pitchers positive credit for good performance (or we measure value relative to league average… but that goes against the whole “absolute value” idea).

Guy,

There’s the easy WPA calculation, and then the very complex one.  The easy one assumes that both teams are equals at all times.

The complex one is what a true gambler would use.  Before the game, you assign a win probability to each team, say .70 for the Twins with Santana as the starter, and .30 for the Royals.  You pre-assign the win values to every player on the roster.  Say you give Santana +.25 wins, you give Nathan +.05 wins on the 50-50 chance that he might come in and, the likelihood that he will perform his usual stuff.  You give Rincon +.00 because he pitched the last 4 days, and he’s not expected to come in today.  You go through the 50 players, and the Twins players get +.20, and the Royals players end up with -.20.

Then, as the game proceeds, using the actual batting lineup, you start assigning WPA based on the expectation of the exact following batters performing a certain way based on the base/out state, the type of pitcher, and the quality of fielders he’s facing, as well as the “protection”, and what have you.  It is a pure gambler’s viewpoint, because the gambler wants to know in real time all this stuff.  (And managers and fans do, I would hope).

If Santana is injured after the first pitch, he gets -.25 wins for not performing as he was pre-assigned to perform.  You could in fact dock him even more, because he now forces your emergency bullpen guys to come pitch.

This is so complex, that unless someone is paying me loads of money, is not worth doing.

***

As for the realness of Jeter/ARod, or the monkey/me, the monkey has a real million dollars, and the Yankees have real wins and losses, and Jeter has an all-star lineup of babes, and ARod does not.  Does Jeter *deserve* it?  Does the monkey?  Who knows.  All I know is that the monkey was there when they were handing out the money, and Jeter happened to be there when his performance helped the Yanks march towards the win.  That’s what WPA is all about.

I find it completely ridiculous to count as zero the contributions of all players on the losing side, regardless of their performance.  The job of a player is not to guarantee that his team will win the game.  The job of the player is to help in his team trying to win.

Some people work on 100% commissions, and some are paid a certain amount regardless of their performance, and a little extra every time they bring in the business.  WPA is about the latter.  You have dozens of people in your department, each doing their job, trying to bring in the revenue.  The better they perform, regardless of bringing in revenue in the near future, they will be recognized, and compensated.  If the money doesn’t come in, someone’s head will roll, and that’s whoever had the lousy performance in the high LI.

I’ve also read issues people have with WPA, over at BTF and elsewhere. 

For the most part, their issues are a lack of understanding of WPA. 

In other cases, their issue is with the implementation, and not the framework, of the WPA that they’ve seen. 

As I described above, you can have a “gambler’s-proof” implementation of WPA that recognizes all the real contexts in baseball.  So, it’s very important to:
1 - Understand the WPA framework
2 - Separate the objections between the framework, and the particular implementation that is being looked at

WPA is what it is, and tells you the story that it sets out to tell.

"All I know is that ....Jeter happened to be there when his performance helped the Yanks march towards the win.  That’s what WPA is all about.”

I don’t have a problem with crediting Jeter with real accomplishments just because luck played a role (nor does MGL, if I read him correctly). But WPA often does not measure real accomplishments in terms of producing actual runs, or actual wins.  Will a high WPA generally correlate with helping your team win?  Sure.  But so will a high OPS or lnr weights.  And again, WPA is biased toward rewarding players who deliver with men on base, while failing to reward those who get on base at the “right” time.

Just as a good example of a very smart guy making an invalid point (depending on your point of view), Guy said this at BTF

but not give any extra credit to the hitter who had the good sense to get on before Jeter’s hit (his WPA is not affected by what later hitters do)?

Here, he has an issue with the particular implementation of WPA, and not the framework.

Let’s take an even wilder example: how much credit to give the player who gets on base, with Pujols on deck?  In the gambler’s-proof implementation, the Cards have one win probability if Scott Rolen and Pujols is not in the lineup, and another if Scott Rolen is followed by Pujols in the lineup. 

By preassigning Pujols a certain number of WPA (say +.10 wins), you are now free to treat Rolen’s universe as if Pujols is something he can exploit.  If Rolen gets on 1B with a normal hitter behind him, he might get +.04 wins.  With the pitcher batting behind him, he might get +.03 wins.  With Pujols behind him, he gets +.05 wins.  (And if Rolen makes an out with Pujols behind him, he gets a bigger minus than if a pitcher follows Rolen.)

Now, when Pujols comes to bat, he sees a normal runner on 1B, and the Cards are expecting big things out of Pujols.  If Pujols fails to deliver, he gets a bigger minus than if Eckstein was the batter instead. 

And all this makes sense because of the preassigning of WPA.

It really is fascinating stuff, if you jump into the WPA pool.

As long as WPA is not misinterpreted (which it is by many, even smart, people), it boils down to a matter of taste.  The only common thread I see in the supporters and detractors is that indeed it has no practical value, other than as perhaps an MVP-type gauge.

Tango, using leverage to scale value, both in the past and in the future, is utilitatrian and in fact, critical.  It has only been recently that even sabermetricians have been aware that the closers’ value must be doubled because of his average leverage.  However, about the only use I can think of for leverage (other than for strategy deicisions of course) is for relievers.  I assume we can expect the average leverage for any position player to be around 1 at any time in the future. I also assume that we can expect all players, both pitchers and batters, to perform about the same , relative to their own baseline of course, regardless of the leverage.  If that were NOT the case, then something like a WPA metric (with some heavy regression toward the mean) might be valuable for projecting value.

I don’t see anything else to discuss.  Discussing the merits of WPA is a dead-end endeavor.  At least it is for me.  For the same reasons I don’t ever get involved in discussing HOF, All-Star, or MVP credientials (besides the fact that I am not qualified to do so).

Right, the LI for hitters and starting pitchers is roughly 1.0, and should be thought of so going forward.  This link should the LI for all players:

http://www.fangraphs.com/winss.aspx?team=Yankees

Just taking the Jeter/ARod example, ARod and Torre and a million Yankee fans have all said that Alex is “pressing”.  He is in fact not approaching the low-LI situations similarly to the high-LI situations.  (Whether Jeter is Mr. Cool in these situations, or whether it’s good timing, I don’t know.) The WPA of ARod is abysmal (for him) +0.4 wins, while Jeter is at +3.2 wins.  They in fact should be pretty darn close, all timing/pressing issues aside.

Does this mean that Arod will continue to “press” in the future?  Almost certainly not.  “Pressing” is probably transient, not persistent, among players of this stature at this point in their career.  (We’re not talking about Calvin Schiraldi or Andy Benes.)

***

One of the (implicit) components of WPA is LI, and we can see how Ron Villone is used only in mop-up duty, but is performing fairly well.  Farnsworth on the other hand is highly trusted, but not pitching as well.

***

As for the “double” comment for leverage, you probably don’t remember Palmer wrote that 20 years ago in The Hidden Game.  Check it out!

***

Yes, the main purpose of WPA is for MVP-type of discussions.  But, the entire basis of WPA (LWTS, LI, WE) are, separately, each extremely valuable.  WPA is the combination of all that, and is almost certainly the lesser of the sum of its parts.  But still fascinating, especially in real-time.

As I’ve said, it does what it is supposed to do and does it well.  It also incorporates important concepts in sabermetrics.  Yawn.  I could not care any less that A-Rod has performed badly and that Jeter has performed well in high leverage situations (roughly).  The danger in that, is that the next thing you know, people will be saying that a team should rather have Jeter than A-Rod (in the future of course) because, look at what WPA and sabermetricians say about the two of them (of course, WPA in the long run, over many years of data, will have most players right where they are “supposed” to be, based on their context-neutral stats).

I agree about the danger. 

***

Note that we should be doing LWTS by the 24 base/out states, since we all agree that hitters change their approach with men on 1b, men on 3b and less than two outs, etc. 

We should use limited WPA in close games, and the 1b/3b guarding or not the lines, and the OF playing in with men in scoring position, and the likelihood of a steal or sac.

And, LWTS based on the type of hitter/pitcher (GB/FB), and whether the pitcher is pitching from stretch or windup.

To MGL’s point: we should use whatever contexts a player’s skill is in play.  The “basic LWTS”, while doing a good job, doesn’t do a great job of it.  WPA goes completely the other way, and considers every context as if the player’s skill is in play.

What WPA does is in evaluation players is:
1 - forces us to think outside the basic aggregate totals
2 - forces us to appreciate the impact of the reliever at different stages of the game

(Oh, and you could do WPA for relievers, since relievers themselves believe that the crucialness of the situation impacts them.  I heard from a close source of a reliever on a team that did *not* want to pitch when the game was on the line in the 9th inning.  I’m sure he’s not the only one.)

I like WPA a lot.  I just think it is—like ANY metric—important to keep in mind what it does and does not do, and what you are looking at when you consider it.

WPA may not tell us much about a player’s level of actual skill, and it may not tell us much about what WILL happen.  But I find it a very interesting and useful way of telling us about what DID happen.

There may be no such trait or skill that leads to “clutch hitters” (or, they may be rare, and scarcely different from other hitters with similar overall talent), but there absolutely is such a thing as clutch hits.  I really like how WPA shows whether a given player helped or hurt the team’s chances along the way.  It’s the best way I have seen yet to demonstrate when players actually make key contributions (or failures) in important situations. 

It does not imply that those contributions came because of a special skill, and it does not predict whether there will be more such contributions or failures from any given player.  But it shows how much, in a given PA, game, or season, the player has helped or hurt the team’s chances along the way. 

I suspect that by the end of the season, a list of the players with the most WPA will be pretty similar to a more conventional list of the MVP and CYA candidates.  There will be some surprises, too.  That does not mean that the metric does not measure what we thought, nor does it mean that the player is not what we thought.  It means that the player—whether because of luck, small sample size, some level of true ability (be it overall talent or be it “clutch,” if such a thing exists), or whatever—the player had a significant contribution towards team success or failure in that given PA, game, or season.

I’m a little surprised and confused by some of mgl’s thoughts, notably the notion of not wanting to reward a player who gets a single to advance a runner to third, when the run does not score, it never does score, and the team loses.  Did that player not still do his job?  Was his contribution not still a helpful one?  Didn’t he still improve his team’s chances of winning?  Perhaps I’m not understanding mgl’s thoughts about this.

The way I read mgl, it’s “winning is the only thing”, and therefore, past value can only be derived in team wins. 

But, the objective of the player is not to ensure that his team wins.  It’s to perform in a way that gives his team a *chance* to win.  Vlad hit a grand slam in the playoffs a few years ago in the 9th inning, a game in which his team lost to the Redsox.  According to the “winning is the only thing” viewpoint, Vlad may have as well struck out in the 9th inning, since the end result (loss to Angels) gives you the same thing.

That’s not a viewpoint I share.

I’ve posted an article at THT today that I hope shows a new insight we can gain from WPA.

Very interesting piece, Studes. However, I’m less impressed by the .7 R2 for WPA and actual RS or RA, and more struck by how these disparities suggest real problems with WPA as a value measure.  The Dodger hitters are +40 RS, but just .57 WPA; their pitchers are just +15 in RA, but look much better in WPA (1.43).  AR hitter are avg, but WPA says they lost 3 games. PIT’s pitchers have been twice as bad as their hitters (vs. avg) in real RA/RS, yet WPA says the hitters have been twice as bad.  Milwaukee hitters are -20 RS, yet +2 wins in WPA.  Etc..

You feel that Milwaukee hitters are scoring at the “right time”, while Pitt hitters are scoring at the “wrong time.” I don’t see how this can be meaningful at the team level.  Presumably, Milw is scoring more late in games, Pitt a little more early.  Would Milwaukee have won any fewer games if their hitters had scored those runs a couple of innings earlier?  Do teams that score earlier generally win less? (presumably not) And this is purely a function of inning/score, not hitting well with runners on:  Milw hitters are actually underperforming their EQR, while Pitt has scored a few MORE runs than their raw hitting stats would predict. 

I guess this is a somewhat existential question, but personally I just don’t see how we can say these WPA numbers are giving us a better measure of these players’ value than their actual RS and RA.  (And I understand we’re talking only about value, not repeatable skill.) To some extent, these players are determining their own leverage:  the game being close in late innings may be a function of the same hitters’ poor performance earlier in the game, while scoring early then reduces leverage in late innings.  Livan looks a little less bad in WPA because he gets hammered in the 1st inning – I don’t think Frank Robinson (or the Nats bullpen) would tell you that’s a plus for the team.

* * *

I would also note that the total pitching WPA for all teams is +25 wins, while hitters are –24.  That is, the average pitching staff is almost 1 win above average, while avg team of hitters is 1 win below average!  I suppose that is somewhat a reflection of teams’ ability to leverage their best relievers.  But it may also reflect a systemic advantage for pitchers in that they can perform poorly in the 9th inning yet still post a positive WPA, while hitters can never be rewarded for poor performance (but can be penalized for good 9th inning hitting in a losing cause).  A good example is a recent game you highlighted, when AZ beat CO 8-7 after game was tied 1-1 end of 8.  In the 9th, CO pitchers gave up 7 runs and got WPA of –49.8 (1 loss); AZ pitchers were nearly as bad, 6 RA, but got a WPA of +0.2.  (And CO hitters got a –35 for the day, while scoring 7 runs!).

Guy, the premise here is simple and essentially proves itself.  Take the calculated differences between batting and pitching WPA and voila!—you have the pythagorean variance.  Almost exactly.  I don’t know why you say that “this can’t be meaningful at the team level.” What do you mean by meaningful?  If it equals the pythagorean variance, that’s pretty darn meaningful to me.

Next, you say that “I just don’t see how we can say these WPA numbers are giving us a better measure of these players’ value than their actual RS and RA.” but I’m not sure where I said that in the article.  Could you please point that out to me?

I don’t know why pitching WPA doesn’t equal batting WPA, because they are equal for every game.  Every change in leverage applies to both pitchers and hitters.  But I double-checked the numbers David sent me, and that’s what they say.  There may be a player categorization issue.

One possible way to fix some of what I think is wrong with WPA would be to use run expectancy rather than win expectancy.  That credits hitters (and pitchers) for performing well with runners on base, but leaves out the score/inning dimension.  As Tango argues, players do change their approach based on out/runner situation, while that’s much less true for score/inning.

I’m not sure I’m doing the math right, but wouldn’t dividing a player’s WPA by their LI essentially give us that?

Protrade uses run expectancy for their metrics instead of win expectancy.  I don’t know the answer to your math question but I’m totally guessing that the answer is no.  Hey, not knowing never stopped me from answering before.

One last thing, Guy.  I think your interpretation of the difference between RS and batting WPA is incomplete.  For instance, if Pittsburgh scores a bunch of their runs in blowouts, then their batting WPA will be relatively less.  And I’m sure there are many other issues and situations that would drive a difference between the two.

"If it equals the pythagorean variance, that’s pretty darn meaningful to me.”

Mathematically, it HAS to equal the pythagorean variance—no other outcome was possible (since WPA is based on actual wins and losses).  The question is, can WPA tell us whether the hitters or pitchers got lucky/unlucky?  I’m not sure that an answer based on changing win probability at that moment really tells us that (whereas if a team scores more runs than raw hitting stats predict, I think you could say those hitters got lucky).

“Next, you say that “I just don’t see how we can say these WPA numbers are giving us a better measure of these players’ value than their actual RS and RA.” but I’m not sure where I said that in the article.”

You didn’t. Sorry for the confusion. I was refering to others in the thread (and at Primer) who argue that WPA is a better measure of actual value than lnr wts, OPS, etc.  I think your position is that, at most, WPA could be an MVP vote “tie-breaker;” but I’m not convinced it’s even good for that. 

“I don’t know why pitching WPA doesn’t equal batting WPA, because they are equal for every game.  Every change in leverage applies to both pitchers and hitters.”

WPA for hitters and pitchers has to add to zero, but I don’t see why hitters and pitchers, as groups, each have to add to zero.  I don’t think that’s true.

"Take the calculated differences between batting and pitching WPA and voila!—you have the pythagorean variance.  Almost exactly.”

Mathematically, the total calculated diff’s HAD to match the pythag variance, since WPA is based on actual wins/losses.  The question is whether the hitting/pitching split assigns responsibility for the variance in a meaningful way, i.e. does it tell us which half of the team got lucky/unlucky?  I’m not sure that’s really meaningful—it’s TEAM luck to win close games. (Whereas, with a team that has more RS than their lnr wts predicts, you can say the hitters got lucky.)

“Next, you say that “I just don’t see how we can say these WPA numbers are giving us a better measure of these players’ value than their actual RS and RA.” but I’m not sure where I said that in the article.”

You didn’t say that. Sorry. I was responding to others who see WPA as measuring actual value better than context-neutral stats like OPS or lnr wts. 

“I don’t know why pitching WPA doesn’t equal batting WPA, because they are equal for every game.  Every change in leverage applies to both pitchers and hitters.”

Total WPA has to be zero (and it is), but I don’t see why pitchers or hitters have to be zero.

Well, we’ll never see eye-to-eye on the usefulness of WPA.  I think that real-time context is useful even after the fact and you don’t.  Oh well.  That’s cool.

WPA for hitters and pitchers has to add to zero, but I don’t see why hitters and pitchers, as groups, each have to add to zero.  I don’t think that’s true.

If, at the end of every game, batting and pitching WPA cancel out for opposing teams, then season totals should reflect that too, no?

Sorry for double post—missed the page change. 

The pitching cancels out the HITTING of the opposing team, and vice-versa.  So there’s no reason that total pitching can’t be positive. 

Hey, maybe Bill James was right about the 52-48!  (kidding....)

Ah, of course.  Brain cramp.  Sorry.  It would be interesting to see if pitching WPA has a tendency to be positive over time, or if that was just a short-term thing in the first half.

As I think you suggested, it could very well be the result of ace relievers pitching in close games.

Isn’t there a small systemic bias in favor of pitchers?  Pitchers can perform poorly in the 9th inning yet still post a positive WPA (if team wins), but are always rewarded by WPA if they pitch well in the 9th.  Hitters are the opposite:  they are never rewarded for poor 9th inning performance (it always reduces WE), and can receive a negative WPA for good 9th inning hitting in a losing cause. 

As I mentioned, a good example is when AZ beat CO 8-7 after game was tied 1-1 end of 8.  In the 9th, CO pitchers gave up 7 runs and got WPA of –49.8; AZ pitchers were nearly as bad, 6 RA, but got a WPA of +0.2 for the inning.  (And CO hitters got a –35 for the day, while scoring 7 runs!).

The reason that the WPA of all hitters is what it is, is because I produce the WE/LI charts based on 5.0 RPG.  The current league total is 4.9, meaning that the average pitcher will come out as above average, against the 5.0 RPG baseline.

***

As for the hitters over/underperforming the WPA, it would not necessarily be about scoring early or late, but more likely about scoring in bunches.  Piling up runs in a blowout hardly adds to your chances of winning, but certainly pads your regular stats.

***

No, WPA/LI will not give you LWTS.  Take for example a guy who hits a grandslam (say LI of 5), and gets ten batting outs in a blowout (say LI of 0.5 for each).  His WPA will be very high, while his LI will 1.0.

However, doing WPA/LI, and *then* summing will give you what you want.  So:

sum (WPA/LI) = LWTS

sum(WPA)/avg(LI) will not

WPA, by definition, must equal to zero at the league level for hitters, pitchers, runners, fielders, if you use the league average runs per game.

Don’t know.  I’d have to study the results a while to see if there’s something systemic.  As you know, there are many, many variables at play.  One example is just one example.

I do think WPA overvalues relievers vs. starters compared to “true value.”

I don’t necessary ascribe to the theory that “winning is the only thing that counts.” I merely bring it up to illustrate that it is a legitimate criticism of WPA.  More importantly, I bring it up to illustrate that as an MVP (retrospective) evaluation tool, WPA is one of an endless number of reasonable ways to evaluate and reward past performance.  Just because something is “elegant and clean” does not necessarily make it any better that any other measure.  That is what I think the allure of WPA is (it’s elegance). I think it is a false sense of allure, however.  Using Runs and RBI, the simplest of all retrospective and certainly context-dependent stats, is as good or better in my opinion than WPA.  You could even do a combination of WPA and Runs/RBI by weighting the Runs and RBI by the LI when they occured.  What could be better than that?

WPA doesn’t overvalue anything.  It properly values everything within the context presented. 

If there are more WPA wins for offense than defense, then this means that your run per game baseline that was used to generate the WE matrix was too low.  If more WPA wins for defense, then the RPG baseline was too high. 

If it looks like relievers are getting “too much” WPA relative to starters, then you could make the argument that once a reliever enters the game, that the baseline should be changed, since it’s easier for a pitcher to relieve than to start.

If there are more WPA wins for offense than defense, then this means that your run per game baseline that was used to generate the WE matrix was too low.

Tango, if batters bat worse in close and late situations as a group, won’t total batting WPA be lower than total pitching WPA?  Yes, the two will cancel each other out, but I could see this happening.

Good point. 

I think the answer is “mostly no”, because if batters are performing at 4.0 RPG in close & late situations, and if there are 10% of those, then they must be performing at 5.11 RPG in the other 90% situations.  Assuming the LI is 2.0 in those 10% situations, then their RPGxLIxPA in this case comes out to
= 4.00 x .10 x 2.0
+ 5.11 x .90 x 0.89
= 4.89

I doubt it’s anywhere close to that.  If anything, it might be 4.90 in high LI and 5.011 otherwise, giving us:
= 4.90 x .10 x 2.0
+ 5.01 x .90 x 0.89
= 4.99

Which is essentially equal to the 5.00.

Guy reported that each team’s offense, defense was off by around 25 wins per 30 teams, or .83 wins per team over those 88 games, or .0095 wins per game.

Since Fangraphs is using a baseline of 5.00 RPG instead of 4.90 RPG, he is undercreding the offense by .10 runs per game, with an 11:1 runs to win converter, that .10 runs per game is .0091 wins per game.

As you can see, pretty much the entire gap can be explained by the baseline.

(Or, if you want to work backwards, the hitters are being shortchanged .10 runs per game x 88 games x 30 teams = 264 runs, which you divide by around 11 to get 24 wins.)

Interesting.  The exact figure, FYI, is 24.87.

Don’t you have to multiply your figure by two to get WPA?  Would that imply that half the difference is your table?

No. 

The WPA figure is wins above average.  If you see Pujols with 550 “WPA points”, that means he has +5.50 wins above average. 

For example, the Tigers WPA is 239.1 + 878.7 + 382.2 which equals 1500.0 points, which s +15.00 wins above average.  Their W/L record is 48/19, meaning that an average team would have gone 33.5/33.5, making the Tigers +14.50 wins above average.  (I guess Fangraphs is missing a game where the Tigers won somewhere.)

Over at BTF, there are a couple of posters that are on the right track.  I recommend to everyone to read this:

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

You can skip down to where I have an image.  WPA is equal to Win Advancements minus Loss Advancements.  My preference is in fact to keep those two separate, and let the reader decide whether to subtract them (as WPA does) or create a percentage (as Mills Brothers do in PWA).

I did this WA, LA for 1999-2002, so perhaps I should release that file at some point.  The most interesting player on the list was John Wasdin being the worst pitcher, even though he pitched ok overall (he was horrible in the clutch), and while Bonds was the top hitter, he wasn’t nearly so outworldly (as his walks likely were issued when they didn’t do as much damage as otherwise).

Ah, games vs. wins.  Sounds like the recent SABR-L debate.

Don’t look now, but Barry Bonds is now #2 among non-pitchers in WPA in the NL, behind the runaway leader Pujols.  And these two guys have missed a bunch of games.

***

studes, there’s really nothing to “debate”.

studes, there’s really nothing to “debate”.

Nice to see the self-confidence, Tango.  Stay away from those SABR-L chats! 

Not so much self-confidence (or arrogance as others might see it), as debating a definition.

I’m just being flip.  I assume you don’t read the SABR-L messages, where there’s been quite the brouhaha lately about how a team that is, say 30-20 isn’t ten “games” over .500, but only five “games” over .500.  Some posters feel that it’s okay to say the team is ten wins over .500 and five games over .500, but not ten games over .500.

In a way, it’s much ado about nothing, though the nomenclature is relevant for your example.  Still, I understand the math and what you’re saying.  No argument there.

When you read SABR-L, you realize that anything is debatable.

Right, that’s really meaningless, and not worth debating.  Once I define a “game” as being W-L or (W-L)/2, then who cares.

In my particular case, we are talking about wins above average, so it is (W-L)/2.

Studes said this at BTF:

Bottom line, if a team has played ten games and everyone on a team has 0 WPA and one person has 10, that team will be 10-0. If he has 8, the team will be 9-1, etc.

False.

If Pujols has a WPA of +10.0 wins, and the rest of the team is 0, then he turns a 10-10 team into a 20-0 team, or an 81-81 team into a 91-71 team.

Remember, WPA = (W-L)/2 .  So, 2*WPA = W-L

The above is based on the WPA definition of starting the game at .500, as fangraphs does.

If you start the WPA calculation with the game at 0 (like in my link earlier), and either side going to 1 win and 1 loss, then WPA = W-L, and studes is correct.

Guy discusses the Cardinals game at BTF, and says this:

Pitcher IP, ER, WPA
Hancock 2.0 0 29.8
Wainwrt 1.2 0 23.5
Marquis 8.0 2 22.4
BLooper 1.0 0 14.9
Isringh 1.0 0 14.9
Johnson 0.1 0 6.3

As I’ve stated, it’s better to keep the win points and loss points separate, since there is value in just going out there.

Let’s say the above is really this:

Pitcher IP, ER, WPA, Plus, Minus
Hancock 2.0 0 29.8 40 10.2
Wainwrt 1.2 0 23.5 50 26.5
Marquis 8.0 2 22.4 60 37.6
BLooper 1.0 0 14.9 20 05.1
Isringh 1.0 0 14.9 35 10.1
Johnson 0.1 0 6.3 06 00.0

So, that represents their involvement during the game. 

(Guy later talks about future prediction, and in no way should anyone argue that we are talking about forecasting.  We are simply doing an accounting of the game, like Bill Gates’s bank account growing faster than Steve Jobs.  But, put them both in a new company, and maybe Jobs does better.)

Anyway, by breaking it up by the Plus and Minus (Wins and Losses), now the reader is free to adjust to his heart’s content.  He might figure that a “replacement” level pitcher would have had 40 win points and 57.6 loss points, if he was in Marquis’ shoes.  That puts Marquis at +0.42 wins above replacement.  A replacement pitcher in Wainwright’s shoes might have gone 35 and 41.5, giving Wainwright +0.30 wins above replacement.

Take WPA for what it is…

Tango: I only raised future prediction as a thought experiment, to get people to think about what’s valuable:  if you knew that these pitchers could replicate this peformance every 5 days (obviously unrealistic), which pitcher would you want on your team?  Which would help a team win more games?  I’m quite clear that WPA is a measure of value, not ability.  Also, I’m responding to some people there who, shall we say, are less nuanced than you in their appreciation of WPA. (And I should say, I think WPA is a neat and useful way to think about real-time games.)

* * *

What do you think of the idea I posted there of converting WPA to total wins, instead of wins-above-average?

From that “thought experiment” standpoint, I agree with you.

***

As for converting WPA to wins:

For background, WPA itself is the difference between win accumulation and loss accumulation.  In the typical game, the winning team will accumulate around 0.90 wins and 0.40 losses, while the losing team will naturally do the opposite (0.40 wins and 0.90 losses).  The difference in win accumulation and loss accumulation is always +0.5 wins for the winning team and -0.5 wins for the losing team.

What you are suggesting is spreading the .500 starting point to get “absolute” wins, and (of course, you’d have to spread the .500 to get “absolute” losses as well).

As long as you can live with crazy numbers, like wins greater than 1 or negative losses, etc, then it’s fine.  I’ve done this in the past, and didn’t really like what it gave me.

I prefer the win and loss accumulations, since it’s much clearer as to what it’s doing.

Tango: I can see why W and L is more intuitive than “negative wins”, though I imagine we’d often net them out anyway. 

So, if you wanted to do this in absolute terms (not repl level), then you would distribute 12.5 W points and 12.5 loss points per game among the pitchers, based on IP, then add WPA?  And do the same for hitters, based on PA?

You’d have to distribute the 50.0 win points to the players involved, which means to distribute 25.0 win points to the offense, and 25.0 win points to the defense.

Since the defense includes fielders, you’ll end up giving 30.0 win points to the nonpitchers and 20.0 win points to the pitchers (of course, with a Nolan Ryan-type pitcher, the fielders have a smaller share).

So, with 30.0 win points to the nonpitchers, each player would get around 3.0 win points.  (Distributed, probably, based on PA for 25.0 win points, and the other 5.0 win points based on balls that are “fieldable, but not routine”.)

The pitchers get 20.0 wins, which you can distribute based on “innings”, though again, you have to be careful with the Nolan Ryan / Brad Radke situation.  Basically, giving each pitcher 2.2 win points per inning.

And you’ve gotta hand out loss points as well.

Actually, because of the way you split offense/defense, and hitting/fielding/pitching, it really becomes a nightmare.  Off/Def is 50/50, and pitching/fielding is about 70/30, but non-pitchers / pitchers is 60/40.  I know it doesn’t sound right, but that’s what it is.

I strongly recommend staying away from distibuting the fixed win and loss points.

False.

Very discouraging.  I shall go hide in a cave and ponder this obviously complex concept!

Take your head out of the oven, Studes.  I think Tango has this one wrong.  WPA = W-L, not (W-L)/2.  For example, the Red Sox are 950 WPA, or +19 (before tonight’s loss), and their record is 53-34.  So if the Cards are 10-0, their players will indeed have a total WPA of +10 (500 WPA). 

And to compare WPA to wins in the sense of something like WARP, where we assume about 10 runs equals 1 win, then you have to cut WPA in half.  Pujols is about +12 wins in WPA, but that’s parallel to about 6 WAA or 60 RAA (leaving LI aside).

Or, maybe Tango was thinking 100 WPA = 1 win, while Studes was using a 50 WPA = 1 win conversion (which is how I think about it), and you were just talking past each other.

Fangraphs is starting at 50.0 win percent, and is going to 100.0 win percent or 0.0 win percent.

So, when Fangraphs has the Redsox at 950 “points”, that’s exactly like saying “950 win percentages points”, or +9.50 wins.  It is simply unfortunate that Fangraphs is starting the scale at 50.0, instead of the clearer .500.  I’ve recommended to them that they convert to the .500 scale.

Pujols having 600 points means that he is +6.00 wins above the average hitter.

Therefore, that means he turns an 81-81 team into an 87-75 team.

Fangraphs_WPA / 100 = (W-L)/2

Normal_WPA = (W-L)/2

Right: you were saying WPA/100 = (w-L)/2
Studes was saying WPA/50 = W-L
So you were saying the same thing.

Can someone please post the BTF link as I can’t find it. That site is so damn hard to navigate round ...

John, click my name.

And I support Guy’s recent post, in correcting Mike’s position.

Hmmm… try this link

Thanks for the link Tango - there was quite a lot of stuff to get through.

I have come to late to this debate by virtue on being on vacation for the past week and a half so rather than regurgitate many of the arguments I’ll just briefly chip in my 2 cents worth.

Before I begin let me say that I sit in the camp that welcomes WPA - for me it is a valuable, but certainly not flawless, tool for certain situations.

First, I don’t hear many analysts making the argument that WPA has any predictive value. As MGL correctly points out it may have some MVP merit and it does tell us which players performed well when the game was on the line, but we all know that clutch skill, if it exists at all, is tiny, and as WPA measures largely situational contexts its predictive value will be low.

Where WPA is interesting for me is in the real-time change in odds of winning a game given the GAME SPECIFIC context. It sort of reinforces what you think in your head as you watch a game (eg, a lead-off HR in the first is nothing to get too excited about, but it is in the nineth, provided the game is tied / close). This is why I think it would be useful to see WPA (and LI for that matter) in real time. Also looking back on the game WPA charts the story of the individual events which impact the game.

Beyond this game specific context, for me at least, WPA has limited interest.

I liked the THT piece on team WPA and variance from Pythag but I agree with Guy’s point. By definition the variance to Pythag is explained by WPA. It is like saying the variance in Pythag is explained by W-L record, because W-L can be derived directly from WPA. I guess where it gets a little more interesting is to break it down to the next level and understand what parts of the team have contributed positively / negatively to the variance.

The other general point I have on WPA is on the picthing side. Surely it isn’t fair that fielders do not get creditied with WPA for good / bad fielding performance. If in a bases loaded 2 out situation the SS fumbles a routine groundball the pitcher gets debited the WPA (this is at least how I understand it, please please correct me if I am wrong), but this shouldn’t be so. Also, to properly look at WPA for pitchers we should also regress those elements of their performance which are more than likely down to luck - for example, an extraordinarily low BABIP - but regressing WPA data would start to defeat the purpose of using it (because in a game cumulative WPA wouldn’t balance). Otherwise it has the same issues that pitcher W-L record has in assigning wins to picthers.

Again it reinforces my top point, we can use it to paint the game specific situation, but its value in any other capacity is more limited.

I think we all need to separate our observations about WPA from the Fangraphs implementation.  There is no “one way” to implement WPA.  BPRo does it differently from Rhoids (do they still do it?), which is different from BPro, which is different from Tango which is different from Jim Albert, etc.

It so happens that Fangraphs isn’t allocating WPA to fielders because that is really hard to program.  Probably impossible unless you log a game.  However, people who use my spreadsheet to log games often put in a fielding/pitching split (there is a special cell to do that).

I disagree with you and Guy about the usefulness of the Pythag/WPA analysis, but then again I wrote it, so I would.  I think WPA is a great use of Pythag analysis, because one-run games are often best dynamically analyzed in real-time.  If a team goes ahead early and the bullpen almost blows the lead, but doesn’t quite, WPA captures that dynamic.  It also captures the bullpen holding the opposition scoreless for the last three innings, and it values the come-from-behind offense that just grabs the one-run game.

I often hear people assume that a good record in one-run games implies a strong bullpen.  That may be a managerial strategy that makes sense, but it often isn’t a reflection of what has really happened in a game or season.  WPA captures well what really happened, in this instance.  IMO.

Of course, you can take the stance that Guy does and note the one-run games are a team’s responsibility and shouldn’t be broken down further (I’m probably paraphrasing Guy incorrectly).  I can understand that, but I do think there is also value in breaking it down.

I echo studes on the fielding.  The WPA framework absolutely allows for fielder allocation. 

It all depends on how you want to implement it.  Fangraphs chooses not to, because it is harder to program.

I was watching a game the other day, and the graphic was Jose Contreras is 17-0 in his last 17 decisions.  No one in the world thinks that the 17-0 is indicative of his true talent.  But, it is a snapshot as to what he’s been involved in.

That’s all WPA is.  A numerical description of your emotions as you watch a game.

Studes:
The fact that your total WPA-pythag differentials matches the actual pythag differential really is tautological; it can’t validate the analysis.  That said, the analysis could still be meaningful—I have an open, if skeptical, mind.  In essence, when a team wins a game 4-3 you’re looking to WPA to tell us whether the hitters deserve more credit for scoring “just enough” runs that day, or the pitchers for suppressing opposition scoring exactly the amount needed to win.  Looking at the game retrospectively, it seems a rather odd question.  WPA will give you a real-time based answer, but is it really telling us something important?

One thing to look at is whether a team generally scored or prevented runs in an efficient way. For an offense, I believe that generally means a small SD—consistently score as close to the mean as possible.  For defense, it’s the opposite:  get torched a few times and give up 17 runs, then give up 3 runs or less in as many games as possible (large SD).  So, you might look to see if this pattern matches up to your WPA findings—e.g. does the Brewers’ offense have a low SD, the Pirates a high SD?  If so, it becomes more likely that WPA-predWPA is really telling us that a group of hitters or pitchers performed especially efficiently or inefficiently. 

Alternatively, we may just be seeing the odd, random way that changing leverage affects WPA.  You say that WPA “values the come-from-behind offense that just grabs the one-run game.” Yes, but that late inning leverage may have been created by those very same hitters’ earlier futility, for which WPA did not punish them severely because they had the wisdom to fail in the early innings.  As I scan through individual games at FanGraphs—definitely fun—I find a lot of games where the WPA result is wildly at odds with what any kind of runs above/below avg analysis would tell you.  Perhaps in the aggregate it tells an important story, but I wonder....

Studes,

I don’t disagree with you about using WPA to analyse close game, in fact I think I aluded to than point in my earlier post. The point I was making is the same as Guy’s - the variance between Pythag and actual W-L is by definition accounted for by WPA.

Anyway, I hear you completely about using it to break down how the game ebbed and flowed - especially for tight games. However, there will still be situations where the usefulness of WPA will be limited. Take the extreme example of a high scoring close game that goes into extra innings, where the visiting team gets a run in the top half of every inning and the home team gets a run in the bottom half of every inning. Looking at the WPA picture over the whole game will paint the story; looking at cumulative game end WPA, IMO, is less useful, unless you understand the actual context of the game. But that doesn’t mean that it isn’t an interesting tool to play with and look at post-game—to your point Studes, WPA will tell us that the offense (on both teams) were great, but the pitching (starters, relievers and closers) sucked, but to know how they sucked requires more information.

Just a small point on fielder / pitcher WPA splits. Would the idea be to allocate the entire change in WPA to either the pitcher or fielder depending on the play or would you actually split it (in a similar manner to how Tango outlined the gammbler WPA implementation)? I guess care needs to be taken to ensure that the allocation has the correct baseline.
_______________
“That’s all WPA is.  A numerical description of your emotions as you watch a game.”
_______________

Excellent description, Tango. I concur exactly.

Guy, I agree that having the variances add up isn’t indicative of that.  I should have said much earlier that that was a good point.  My point stil is that WPA reflects well on the give-and-take of close games.  Sounds like you don’t agree, at least intuitively.  I have yet to score or see a close game for which WPA didn’t reflect well on the game dynamic, but this is ultimately a matter of preference and intuition.  BTW, I’ve personally scored maybe 50-60 games using WPA, which is hardly a huge sample size.

John, the “right” way to split fielding and pitching (I think) is to record two plays—one in which the outcome is based on what typically happens to the batted ball, and the other based on what actually did happen.  That takes a ton of data and confidence.  So my spreadsheet, for now, relies on a simple split between fielders and pitchers.

On my blog, we’ve had conversations on the best split between fielding and pitching based on observations, but it’s all conjecture and guesses at this point.  The safest/best thing to do is with errors; enter them as two separate plays: one in which the out is made and credit given to the pitcher and the other in which the batter is safe and full debit is given to the fielder.  Assigning offensive responsibility in the second entry is problematic, so I usually give it to “No one.”

I’ve had a number of posts and discussions on my blog the past two years discussing these sorts of issues.

Someone is tracking the Yankees WPA, and I made a post here.

Over at BTF, Mike makes the following claim:
Late-game performance (7th inning on) declines across every LI situation.

I’ve provided extensive data here:
http://www.tangotiger.net/files/clutchdata.txt

A simple average of the data presented shows a wOBA drop to .343 in the high-pressure situations, from the otherwise .351.  The OBP barely changes to .348 in high-pressure, from the otherwise .349.

I seem to remember concluding that most of the drop was almost entirely due to the relief pitcher.  That is, the .351 to .343 drop in wOBA is almost completely explainable by the quality and role of the opposing pitcher.

In any case, it doesn’t really matter.  Even if I have a more fluid baseline (which includes the HFA, quality/role of opposing pitcher, and the “shift” in performance for the hitter/pitcher based on leverage), etc, the end-results will be very similar.  Jeter will be around +3.2 wins, regardless of how I generate the win probability numbers.  The underlying rule, by definition, is that the average hitter will always end up with a WPA of +0.00.

This link may also be useful to some:
http://www.tangotiger.net/innwin3.html

Tango,

I see that you still you A.S.S to parse data! It is amazing how good that application was / is. There is still nothing quite like it; shame it hasn’t been updated ...

"My point still is that WPA reflects well on the give-and-take of close games.  Sounds like you don’t agree, at least intuitively.”

I don’t disagree that it reflects the give-and-take.  What I’m questioning is whether, once the game is over, those WPA figures really tell us how valuable the performances were (to the extent WPA departs from RS or RA).  On Friday, Carpenter and the Cards beat LA 5-0.  Carpenter got 36 WPA—not a lot for a complete game shutout—because StL took the lead early by scoring 4 runs.  By reducing leverage, the StL hitters minimized the value of Carpenter’s late-inning excellence, and increased their own net WPA by minimizing the negative scores for their failure to score in the later innings.  Now, suppose the StL hitters had scored all 5 runs in the 8th.  Then Carpenter would have had 60+ WPA, while the hitters collectively would have been in the black.  Would that change in timing have made Carpenter’s performance almost twice as valuable?  Why?

I realize that I’m “cherry picking” games where WPA doesn’t match traditional valuation.  But I think that’s fair, since A) this happens quite a bit, and B) the point of your article is precisely that WPA tells us something important when it deviates from what RS/RA tells us.

Guy

I think this is to my point that WPA should only be looked at in *game specific* situations.

For example in the game you outline, Carpenter actually had the highest WPA—almost double the next highest, which was Duncan. For this specific game Carpenter clearly contributed most to the win (in terms of WPA).

From WPA (and the graph) we can tell that Carpenter obviously dominated -never really got in to trouble, and didn’t have to call on the bullpen to bain him out.

However, I don’t think we should be comparing this game to a similar game where the Cards score their runs later. In this case Carpenter’s LI will be higher and WPA = LI*performance, his WPA will be higher. But for a shutout he will still likely be the most valuable player for that game.

So, I think in this case WPA does / can paint the correct game specific picture.

John:
I don’t think we’re really disagreeing (and I don’t think Studes and I are actually that far apart).  My disagreement is with those who feel WPA is a good measure of total value, a metric that might inform your MVP or Cy Young selection for example.

On Friday, Carpenter and the Cards beat LA 5-0.  Carpenter got 36 WPA—not a lot for a complete game shutout—because StL took the lead early by scoring 4 runs.

I get confused posting about this subject with you, Guy. I thought we were talking about close games.

If a starter keeps a game close in a low-scoring game, he gets a lot of credit.  If he “keeps it close” in a high-scoring game (i.e. gives up a lot of runs but the game is close anyway), he doesn’t.  In that example, RS and RA agree.

I agree that WPA is not a very good tool for assessing player value, though I still think it has value as a secondary consideration.  However, I would use it (secondarily) within a “class”: starting pitchers, relievers and batters.  I don’t think it works well across those classes as a value stat.

But “value” is different than assessing who contributed the most to close games.  By definition, RS and RA will get about equal value in close games.  It’s the timing of when RS and RA occur that help make a game close.  Assessing the “responsibility” of close games is as much a timing issue as an absolute value issue, IMO.

Close games intensify late in the innings.  If there is such a thing as “clutch”, it appears in close games.  WPA captures both of those dynamics. 

I know you don’t value those aspects of a game after the fact.  But I think they’re the essence of what makes a close game unique and interesting.

You get a high WPA for:
1 - performing well, and for
2 - being at the right place at the right time

Much like life.  WPA is a bank account.  After 10,000 “trades” (plate appearances), the noise of #2 is washed out by the impact of #1.

If “value” is defined as the money leaders, the most valuable bank account, then WPA is what I would use.  However, the asset value is not WPA.  The asset value is the true talent level.  That may or may not correspond to the bank account.

If a pitcher happens to pitch in blowouts all the time, then his bank account will grow very little.  His #2 above will depress his #1 above.  If a pitcher happens to always pitch in tight games, then, lucky him.  His bank account will grow alot if his pitches well, and shrink dramatically if he blows it.

Such is life for a day trader.

After 10,000 “trades” (plate appearances), the noise of #2 is washed out by the impact of #1.

Except for relievers vs. starters.

I get what you are saying (Rivera has an LI of 2.0, while the starters get 1.0), but I’ll disagree to a point.

Rivera has a high LI because he is Rivera.  But, if there was no such thing as leverage, Rivera would have absolutely been a starter.  His overall LWTS * LI would work out to around the same thing, whether he is a starter or reliever (same thing for Papelbon, which is why there is the discussion of whether to make him a starter).

Theoretically, if you line up your pitchers 1 through 12 in terms of “true talent”, their LWTS * LI (or WPA) should also line up.  While certainly not true for a single team in a given year, it’s most likely true for a career, such that the good starters and good relievers all bubble to the top:

http://www.livewild.org/bb/toppit.html

I had forgotten about that site, Tango.  Thanks.  I need to bookmark it.

I’ll buy your career argument, with an option to change my mind.  It’s an interesting point.  Certainly the innate variability in most reliever’s year-to-year performance will help the career standings, too.