THE BOOK--Playing The Percentages In Baseball

Phil, this is a great study, and I don’t think it’s a coincidence at all.  There is every economic reason for a lemons effect to exist in baseball, plenty of private information specifically.  This is exactly what I found in my free agent studies that you mentioned later in the paper.  The most relevant of these is probably this one, “The Cost of OPP”:

http://www.baseballprospectus.com/article.php?articleid=10883

My findings were that free agents WHO SIGNED MULTI-YEARS deals with new teams underperformed free agents who signed multi-year deals with their old team, specifically in future years.  The test you ran in this paper is similar to the one that MGL ran, where he found no effect for all free agents.  The issue with that is that most free agents sign one-year deals and, as both your study and my studies show, the pronounced lemons effect is clearly relevant for players more than one year past departing from their old teams.  Hence, diluting the effect with players who are free agents for one-year deals is going to hide the effect since the player who signs a one-year deal is not one who the old team has the down-the-line private information about affecting their decision.  I suspect that if you ran the same study as you did where you looked only at players with significant runs created in their careers, you might find a pronounced difference in between players who switched teams and those who did not.

For presentation, you might do better to include an aggregate runs above replacement per player or something like that next to the runs created per game.  It’s tough to quickly see the effect of more aggregate production, and then discern whether it’s due to playing time or superior rate of production.

All in all, fantastic work, and I like it even more because it rings true with all the analysis I’ve done on multi-year deals given to free agents!  I continue to think this is a major issue, and it’s great to see work done on the subject.  The sabermetric outlook of thinking of teams as being run by idiots without knowledge about their field is just wrong.  Sabermetrics is an asset, but the old timey baseball scouts are as good at their work as any other profession where people learn tricks of the trade over time.

Thanks, Matt!  I’d seen some discussion of your study, but haven’t read it and digested it yet.  Will do that today.

As for the metric ... I used RC27 because it tends to amplify small differences, so that you can see whether the two groups are actually different.  But perhaps I should do runs too.  I’d have to do runs per (something) in order to keep the productivity numbers separate from the playing time numbers.

>“the player who signs a one-year deal is not one who the old team has the down-the-line private information about affecting their decision.”

I should read your article first, but are you saying that the one-year deal is the “no information case”?  That a multi-year deal with the old team suggests private information about “goodness”, a deal with a new team suggests the old team has private information about “badness”, and a one-year deal with the old team is neutral?

With respect to runs produced, I meant to come up with a summary statistic that was a counting stat and one that was a rate stat, to distinguish the playing time factor from the production factor.  Without this, the reader can see the change in the rate of production, and then tries to deduce how much that really is by looking at the number of ABs.

For the one-year deals, I’m saying this is a situation with the least private information, because both of our studies show that the biggest utility of private information is determining a player’s ability to produce more than one year after the transaction.  The private information in one-year deals appears pretty small regularly as production is similar in the first year after changing teams to the production of equivalent players who stayed put.

Oh, and for the free agent tables 12 and 13, those *are* limited to players with 1000 RC.  The captions should have included that.  I’ll fix later today.

Matt/3: I get it.  I’ll read your study before commenting further.

I agree - great study!
What I particularly like about it is that I think it clearly shows that there is information GMs have (about their own players) that doesn’t show up in the stats. So, when we say something like:
“GM A is an idiot because he is using phenomenon Y as the basis for a trade, but there is no evidence for phenomenon Y”
what we should really be saying is:
“I have no evidence for Y, but GM A has been successful. While I’d like to chalk it up to luck, it’s quite possible that GM A knows Y to be true based on information I don’t have.”

Y can be any of the controversial topics that come up in SABR (clutch, pitcher BABIP control, team chemistry, etc).

@ mettle/#6

I would think that the private information would be about how a player will progress/regress, not about clutch, BABIP or team chemistry.

Alex/7:
In this case, yes, Phil showed the Y has something to do with regression, but the bigger point is that there may be many things GMs know about their players we don’t. These may be based on being privy to dozens of daily BP at-bats, performance on batting and pitching drills, medical reports, x-rays, psychological profiles, intricate metrics of mechanics and who knows what else GMs use to evaluate their multi-million-dollar investments. I think it would be naive to assume they only use on-field measures.

Good stuff Phil.  It is disturbing that the free agent numbers are the opposite of the pre-FA numbers.  Given the theory (that GM’s know more about their players than other teams’ players), you would expect similar results, in that GM’s would tend to let lemons go in free agency.

Could it be that the effect you found was before the modern era of medicine such that some players worked out during the off-season and other players did not (whereas nowadays almost everyone stays in shape)?

Also, I would like to see the average lines for say the 5 years before the year after the trade or non-trade.  First, Marcel only looks at 3 years.  Maybe the traded players had much worse (MLB) years before those 3 years or even worse years in the minors.  Also, as you say, while the Marcel’s could be the same, the “trend” for those 3 years could be different for each group - control and traded.

And, please, why do you use RC?  I hate that stat.  I see .5 runs per game difference and I think 75 runs per 150 games, wow that is enormous (in lwts runs).  But, we have to divide by 9 or something like that to convert RC to lwts, right?  And even then, isn’t RC computed by figuring what a team of those players would score, which therefore doesn’t use the correct weights for each offensive event?

RC?  Horrible!  I’d seriously rather see OPS…

Isn’t RC just a “toy” to see what it would be like if you had a team of 9 such players?  A player’s (context neutral) value if of course not as if he plays with a team of like players, but as if he plays with a team of average players.

For example, if a player had all HR but no walks or non-HR hits, isn’t his RC going to reflect the fact that his HR are only worth 1 run?  And if he is a high OBP guy with little power, isn’t his RC going to be really low (as compared to if he played on an average team)?

Even if that is wrong, the reason I hate it is the scale. I want to know how many runs/wins a player is better or worse than an average player.  Without you telling us (that .5 RC is equivalent to 7 runs a season), I would have no idea.  If you have to tell everyone what your metric means in terms of wins per something, well…

Sure, OPS doesn’t tell me that either, but at least I know that. When you say a player’s RC is .5 per game better than average or than someone else, it implies that that player is 75 runs per 150 games better!

MGL,

Could absolutely be a matter of some players not staying in shape, and those more likely to be traded.  I’ll add that as a possibility.  That kind of half supports the lemons hypothesis, in the sense that (a) there IS a difference with the traded players, but (b) it’s perhaps not unknown to the acquiring team.

I’ll check into the Marcels for 5 years previous and post the results here.

RC27 vs. other stats ... I guess that’s a matter of taste.  I prefer RC27 as a pure rate stat that doesn’t care at all about playing time.  I suppose runs per 500 PA or whatever would be that too, but have the advantage of being in pure runs ...

I didn’t say a player’s RC was .5 per game better ... I said his RC27 (or RC/G) was .5 better.  Completely different! 

Maybe I’m just Old School.  I’ll add another measure for you newfangled guys. 

BTW, that’s a bit how I feel about wOBA ... hard to convert to runs.  And I know what 6.50 RC27 means, but not what .370 wOBA means (although they’re probably not that far from each other, right?).

At least RC27 “means something”, in the sense that it tells you how many runs 9 of those guys would score!

To get runs from RC27, just multiply by batting outs (AB-H) and divide by 25.5.  It’s like ERA, but use 25.5 instead of 9, and AB-H instead of innings pitched.

Anyway, your point is still well taken, and I’ll add something for runs per 500 PA.

Ok, but to compute RC27, does it use lwts values or something else?  Is it as I said above, that a player with all HR, but no other hits or walks would have very little value (because a team of hims would not score more than one run at a time)?

Using anything other than lwts values is just wrong.  That is why OPS is wrong and that is why wOBA is right.  Any other formula is just an approximation of value, some better than others of course.

If you are going to express value/talent in runs in the first place, don’t you have to use the proper lwts value?  Unless you want to express it as a “toy” and not as the real value of a player (to an unknown or an average team).

Expressing a player’s value as the number of runs scored if a team were made up of all those players is a toy, right?  It has no meaning whatsoever, other than it is in most cases close to the right answer, which is linear weights.

And yes, I am talking about about lwts per 150 games or per game or whatever.  If you want a rate stat that is.

Sometimes I use OPS only because it is a little easier to compute and some people relate to it and are familiar with it more than a more esoteric stat, like wOBA. But if I am doing serious research and a player’s value or even relative value is important, I am always going to use wOBA or lwts.  I can’t imagine using anything else.

Then again, is it not that big a deal.

I’ve updated the study.  Same link.

And, er, sorry, I still haven’t gone through Matt’s free agent study.  Guy’s comments at my post kept me busy all weekend.

Oh, sure, blame me!

Small point I mentioned earlier:  I’m not sure it’s fair to describe these players as “traded,” even as a shorthand description.  I haven’t looked at all of the players, but my guess is at least a third and maybe half were not traded at all.  I’m not trying to nitpick—it matters because “trade” implies some other team gave up something of value to get the player, which is an essential part of the “lemon” hypothesis.  In many of these cases, that didn’t happen.

Right, not all were traded traded.  But sold for cash is as lemons-compatible as traded for players. 

As for released players ... I think the lemons hypothesis only requires that the players on the market are worse than public information would suggest.  Released players are players on the market, so they’d qualify too, no?

If I abandon a nice-looking car at the side of the road, there’s probably something wrong with it, even if nobody pays me for it.

Can you summarize the changes?  I don’t really want to read the whole thing again.

Sure.

The ending changed.  There’s now an alternative hypothesis: that the traded players did exactly what they were expected to do—they had fewer AB than the control group because, unlike the control group, they were already on the decline.

That is: the Marcels didn’t take into account that the traded players were decaying fast, so they overestimated their future. 

I created a control group using something other than Marcels (controlling for the previous two seasons separately), and the effect diminished.  But the traded players were still declining despite the attempt to control for that, so that’s probably what’s causing the difference.

"That is: the Marcels didn’t take into account that the traded players were decaying fast, so they overestimated their future.”

Marcel’s or any other projection system that I know of, don’t take that into consideration because there is no way to know that, as far as I know. 

Now, I am not sure if there has been a lot of research on identifying “trends” in historical stats in order to better predict aging curves.

Plus, some projection systems do use different aging curves for different types of players (e.g., fast players tend to age better than slow players, I think).

So, the theory is still that somehow teams know that a player is more on the decline that his age would otherwise suggest, right?

RC27 vs WOBA:

Runs created works as a general measure about as well as the other run estimators, except at an extreme theoretical example like a player who only hits homeruns (even Jose Bautista has more singles than homers, but barely).  Using RC or EQR or LW or whatever is not going to change the results of an aggregate study like this.

For me RC27 has a lot of familiarity to me.  If I were just getting into sabermetrics right now I wouldn’t bother with it and use WOBA instead.  But I’ve used it for many years.  My APBA game ranks players by RC or RC27.  I’ve become accustomed to it as a frame of reference.  I know my 1B better create 6.0 RPG or else I should be looking to replace him, but if my shortstop with a good glove is only around 4.0, he’s worth keeping.