THE BOOK--Playing The Percentages In Baseball

At some point of the season recent stats should trump sample size.  Marcel suggests it would be some time in late August (130 games or so).

Sure. I still don’t want anyone quoting me this year’s stats only, and the exclusion of prior years’.

One of the problems with this approach is that it allows someone to cherry pick stats to support their argument.

For example, in the roundtable discussion I linked to above, one of the guys does not like Volquez (as a top pitcher) because he “has been doing great for only half a season (which is true).”

Yet, the same guy, and all the others, see no problem in quoting this year’s stats only for other players whom they like.

That is what I said, “You can’t have your cake and eat it too.”

I have said this a million times and I’ll say it a million times again, “There is no magic to ‘this year’s’ stats!” They are a little closer in recency to last year’s of course, but last year’s are also part of a player’s sample of performance, as is the year before’s, etc.

You want to talk about how good a player IS, or anything about whom you would “want” in an upcoming game, don’t come to me with anything less than his last three year’s of stats PLUS a regression.  I won’t listen to you otherwise.  And you, as discerning readers or listener’s shouldn’t either.

I also don’t want to hear, “So and so has done this lately, but small sample size caution is in effect.” If you want to tell me “small sample size caution is in effect,” then you had better AT LEAST quote me all of the player’s stats (again, at least last 3 years, if they are available), and do a regression.  Then, and only then, can you tel me about the sample size affecting the reliability of that assessment.  Don’t tell me that Volquez or Lee are great pitchers, with small sample size caution, when neither one of them has been great before this year. Don’t tell me that something is wrong with Sabathia during the first month of the season, but small sample caution is in effect (there ain’t no small sample - the guys been pitching for over 5 years in case no one noticed).  And please don’t tell me that Sabathis is great again, because he has pitched great AFTER the first 2 weeks (as if the first 2 weeks don’t count anymore).

That is another bugaboo of mine.  Ignoring periods of performance as if they don’t exist or otherwise discounting them, like Sabathia’s first 2 weeks.

Or the ubiquitous, “So-and-so played terrible the first month or two months, but is playing great now (what, right this very moment as we speak at 2 in the AM?),” as if we don’t care that much about that first month or two.  After all, that’s over with, it’s in the past, don’t worry about it, it’ll never happen again…

All good points, I didn’t even realize we were doing that.  I think when it comes to Sabathia we jumped to this season’s numbers because it’s just assumed (and shouldn’t necessarily be without evidence) that because his past work has been very stellar that his early season woes were more flukey than indicative of a decline in skill level...almost as if we expect that by season’s end those two early starts will be trumped by the 30+ other good starts… and because of that we forget to note his past work, since the false assumption is made that everyone knows about said past work.

Ignoring those two April starts would be incorrect, but weighting them less would not be, and taking a look at his performance indicators all told: 3.24 FIP, 9.06 K/9, 3.57 K/BB, .261 BAA, 1.28 WHIP, 0.93 HR/9 (compared to 0.75-0.87 in the last three years) would show he is still very good but his barometers make him look worse.

Perhaps not as good as his projection would point him out to be, and perhaps this should have been part of the response to the question.

Presuming that the Marcel weights are correct, then if you’ve got 2.8 months worth of performance this year, that should be as predictive as 3.5 months worth of performance in 2007.

Very interesting request by MGL.  What if you put it up for a vote by the readers.  Would you rather have the whole season of 2007 or only the first 2 months of 2008, to predict what will happen for the rest of the season?  How about 3 months?  4?  5?  Marcel would say 4.8.

Let me run a poll in the morning, and then we’ll have the competition…

but weighting them less would not be...

Less than what?  It would be incorrect to weight them any less than any other 2-week period, other than the fact the first 2 weeks get weighted slightly less than the next 2 weeks, which get slightly less than the next 2 weeks, etc.

No offense to you Eric (you are doing fantastic work), but personally, I don’t want to hear how someone did in any 2 week or 2 month period, no matter how good or bad it was compared to his career or any other period.  I don’t want to know how he has done this year, this past month, the last 2 weeks, etc.  Just give me a player’s Marcel.  That’s all.  Anything else and I am going to make a mistake in assessing or projecting him.

Tango, asking people which is more predictive is different than writing an article and quoting last year’s stats every time you say something about a particular player, with no explanation.  People would think you lost your mind, even though that would be more “correct” than posting this year’s stats, and it’s (apparently) not even close, according to your 2.8, 3.5 month comparison.

For example, “Kenji Johjima?  Why are they even playing this guy?  He is horrible!  .287/.322/.432”

I could post dozens of similar examples, whereby very few people would disagree with the sentiment, but think you were nuts if you posted last year’s stats to “support” that sentiment (the stats would obviously NOT support the sentiment).

It still puzzles me that so many baseball writers (including those with sabermetric inclinations) are quick to judge players based on the less than half a season’s worth of games that have been played so far. People were kicking Jim Edmonds while he was down in San Diego only to sing his praises as he helped the Cubs continue their NL dominance (granted, part of that was because of Edmonds’ injury record and age, but I still think it was too early to write him off). Jason Giambi is another example of this premature judgment, and Nick Swisher is slowly getting back to his usual level of production.

Maybe part of the problem is when people look at projections and expect players to perform that way from the beginning of the season to the end with great consistency, when players often have many peaks and valleys over the course of the season but end up at their level of true talent when the season is over. Or maybe writers just need something to publish and a player’s recent performance is an easy target for an article/post.

At some point, you have to admit that not everyone is writing with the same singular purpose that you are in every situation.  I don’t even disagree with your premise, but the assertion that no one should ever talk about current year performances or small sample performances is a little totalitarian, no?

For instance, my morning fangraphs post for Tuesday talks about Jose Guillen’s performance in his last 156 plate appearances.  I use a small sample that includes random endpoints, so you’re really going to love it! But I find it interesting that Jose Guillen has managed to go 37 games without a walk and hit .360 with a .600 slugging percentage while doing so. 

Is it predictive? I don’t care.  It’s interesting. 

I didn’t read this particular round table, so maybe the critique works in relation to the linked article.  But overall, the desire to see nothing written that deals with the reality of current performance is a desire to see a lot of interesting writing go away.  And I’m not in favor of having less interesting writing floating around.

Tango wrote recently about discounting past performace something like .9998 for each day in the past, but I thougth days were hard to compute, and wasn’t sure how to handle offseason.

Instead what I do is weight current season stats at 1.00, season -1 at 0.80, season -2 at .64, etc, multiplying .8 for each additional season, add them all up, MLEs as well as actual major league performance, and then regress as necessary on the adjusted career total.

Even when weighting the current season at 1.00, 200 PA is still a lot less than .80 * 700 (560) for last season which then I would agree in general with MGL’s comments

I actually had some of the same impressions while reading that post—wasn’t sure whether people were talking about “best” vs. “best so far this year.”

The Sabathia question is an interesting one to me.  I think many of us would intuitively assume that there was some “scouting” sort of thing (mechanics were out of sorts, minor injury, tipping pitches, etc. etc.) going on in his first two April starts, so we tend to discount them.  I know I do.

Still, if we take the long view (as MGL is suggesting) then it doesn’t really matter whether we discount those two starts or not, cause they just don’t have that much impact anyway.

The study here would be quite easy, isn’t it?

Look for star players who had a one-month (or two-start or whatever) consecutive time period where they were abysmal.  It has to be surrounded by great performance (overall) in the one year period directly preceding it and the two months directly following it.

What happens to such a player in the one year following those two months?

Is it simply a situation where it was a short-term situation where mechanics or some real but short-term personal situation kicked in, but was quickly addressed?  And that this player is not expected to be subjected to the same issues that could be back into play?  Or, is it that he is susceptible to again losing it?  Or, is it simply a case of two games (or one month) that happened to occur back-to-back?

We all know the landscape now.  Time for some answers…

David, #7, I am being a little hyperbolic (and sarcastic) in my comments, just to make a point.  I agree that lots of things about a player’s recent performance are interesting and worthy of mention.  As I said, I just don’t want to see it used to support an argument about how good or bad someone IS (as opposed to WAS), or an argument or discussion that implies a player projection.

It is ubiquitous, no?  And given the choice between this year’s stats and last year’s as support for any of those types of arguments, I’ll take last year’s (and I’ll win).

I did a study on a team level last year to determine the optimal number of games to “look back”. There’s a tradeoff, because as you go back further (ie to last season), you increase the sample size but decrease the relevance of the data. The methodology was: if the Mudville Nine have played .560 ball over their last 40 games, then the predicted winning percentage for game 41 is .560. If they win game 41, the error is (1-.56)^2 = .1936. If they lose game 41, the error is (1-.44)^2 = .3136. Repeat the same for all games and all teams during the study period, and determine the number of “lookback” games that minimizes the sum of the squared errors. The study covered every “lookback” segment of games from 1 to 500, and included every team from 1900 to 2006. Some of the interesting results:

- The optimal number of games to look back was 201, although all lengths from about 185 to 230 were generally just as good.

- Looking back 500 games (ie more than three seasons worth) was more accurate than any length under 120 games

I also tried weighting the games in the lookback period by x^y, where y is either the number of games back or the number of days back. It is more accurate to use days back, and the ideal x was .9970 (ie the weight is .9970^days back). A bit different than Tango’s base, but teams are a bit different than players.

It seems like it’s relevant to link to Sal Baxamusa’s article on rolling Marcel projections:
http://www.hardballtimes.com/main/article/the-running-of-the-monkeys/
That’s probably the best available mechanism by which one can use 2008 stats as part of your analysis of player talent...at least among those that I’ve seen.

I’d love to see a site like THT or B-Ref formally incorporate his methodology (or Tango’s, who has a similar approach as I recall) into a line on their player pages.  It would be extremely helpful to see that during the season.

It does look like B-Ref has (with subscription) an “Over past 365 days” split, which is about as close as we can get right now.  I’d have to fiddle around and see if I can get it to post all players’ stats on one sheet for easy copy ‘n pasting.
-j

jinaz, I have a spreadsheet that allows you to get an instant Marcel by copying and pasting data straight from THT.  It should be distributed in a day or two - keep your eyes on THT Live.  It’s a quickndirty tool, in that it uses year-by-year data instead of day-by-day data, but the differences are really pretty small.

We’re also looking into the feasibility of presenting this on our player cards - it may or may not be possible.

Sounds great, Sal, I’m looking forward to seeing it! -j

Dackle, good stuff.

For hitters, I use .9994 and pitchers is .9990

Interesting that for teams you are using .997, which just goes to show you how much randomness there is in team win%.

If you put all those numbers to the power of 365.25, you get:
hitters: .80
pitchers: .69
teams: .33

When you said this:
“- Looking back 500 games (ie more than three seasons worth) was more accurate than any length under 120 games “

I was hoping to see something like: last 120 games = games Tminus121 to games Tminus620

Is this something you can do, filling in “Tminus620” with the appropriate daysago to make this the same.

Of course, at the team level, you get lots of movement which kinda kills it overall.  But, this is the basis of what I’m talking about for individual players.

Maybe this is covered in the Marcel projections (which I don’t know anything about-is there a link which explains the basics?) but couldn’t we simply look at, say, post all-star performance in 2007 (or post June 30) and see whether it correlates more with first half 2007 stats or the entire year of 2006 stats? Isn’t that the basic question?

Click on my name for all things Marcel.

Per this discussion, and Jinaz’s comment: I’ve added to the FanGraphs leaderboards the function to filter on the past calendar year, 2 calendar years, or 3 calendar years.

You can do this for any particular year too, so if you looked at stats from 1990, it would be for 1988 - 1990 if you chose the 3 year option.

It might be a tad slow loading the data, but it’s the best I could do right now.

http://www.fangraphs.com/leaders.aspx?pos=all&stats=pit&lg=all&qual=y&type=0&season=2008&month=10

Tango, it might be a bit tricky to code, especially because it’d be nice to do a cross-section rather than optimize an arbitrary number like 120. 

Not really surprising that the weights drop off faster for teams. The Mariners once played .600+ ball over a four-year period, but are probably not at that level now, whereas individual players (eg Ichiro) certainly don’t drop off that sharply.

Using 200 games back makes the Rays a .475 team (95-105), and even if they play at .475 for the rest of the year that would still put them at 85-77, which is plausible. And of course if they keep winning then that .475 will rise.

David/19: that is fantastically beautiful.

I think the idea of a real-time, daily-updated Marcel forecast would be awesome and I hope a site like B-Ref, Fangraphs, or THT is willing to run with it.

I think the idea of a real-time, daily-updated Marcel forecast would be awesome and I hope a site like B-Ref, Fangraphs, or THT is willing to run with it.

The problem with that (and I am being facetious and serious at the same time) is that it will immediately and abruptly put an end to 95% of all the articles and discussions about who is good, bad, who should be playing, benched, traded, raised or lowered in the batting order, released, who was a good acquisition, bad acquisition, who is a good GM, bad GM, etc., in the newspapers, TV, radio, and the internet.  (Which is a good thing, IMO.) Or at least it should.

I think Fangraphs’s “last calendar year” or “last 2 calendar years” would be almost as good as a Marcel.  When I checked yesterday, CC was #1 in WPA for starters, even given the horrible early season performance.  The only reason we even remember the horrible early season performance is because it was early season.

If we did like what they do in Tennis and Golf, and use rankings of players in terms of NOT resetting to zero at the start of each year, we wouldn’t have all these conversations (just as MGL is saying).

But, spring is hope where everything starts from scratch, and spring is baseball.

I strongly agree with 13 Jinaz and 22 Sky.  And if it’s brought into being, please publicize it so that I don’t miss it. =)

By the way, even just providing the rolling Marcel GPA (environment-adjusted GPA+?) or wOBA or something, rather than making the space to add a whole stat line, would be really keen.

What about this (link)?  (I did my senior thesis on Bayesian probability, so it’s a sweet spot for me - and it seems there should be a pretty great application for it in these discussions)

water: you missed our thread on the subject here:
http://www.insidethebook.com/ee/index.php/site/comments/chipper_does_not_compute/

I did the Bayes analysis in posts 25, 35, 50.

Check it out, and feel free to add your thoughts on that thread.

Is there a missing link in 27 above?

You can click his name.  However, it is Nate’s article on Chipper, which is why I am pointing discussion to that, there.

Sal has a great new article up at THT that uses in-season numbers to project the true talent here: (click name).

Great stuff.  If anything, this will help fantasy baseballers…

I strongly agree with 13 Jinaz and 22 Sky.  And if it’s brought into being, please publicize it so that I don’t miss it. =)

Sal provides some spreadsheets to calculate up-to-date Marcels in his article here:
http://www.hardballtimes.com/main/article/is-this-guy-for-real/

I’m not sure where else to put this, but Sal’s version of Marcel appeared in an article in the Cincinnati Enquirer today to try to understand what (if anything) we can take from Chris Dickerson’s performance.
http://news.cincinnati.com/apps/pbcs.dll/article?AID=/20080906/SPT04/809060374/1071
It’s perhaps not the best use of the system...Dickerson’s minor league numbers tell us that his performance with the Reds this year is way out of line with his talent. 

But still, Marcel mentioned in the paper…
-j

Yes, that is a nice article which discusses the whole issue of estimating a player’s true talent quite well, without going into too much detail.

Dickerson has quite a bit of AA and AAA stats and he has been consistently between an 87 and 94 OPS+ in MLE, which is not bad for a young good defensive CF’er.  OTOH, those minor league numbers are going to knock down that Marcel quite a bit, as the article suggests.  Probably down to a slightly above-average MLB CF’er.