THE BOOK--Playing The Percentages In Baseball

Out of curiousity, what is it you don’t like about Tampa Bay Rays manager Joe Maddon?

It would be fun to try some overs/unders on some of these--trying to pick the instances where the original projections might have been off, due to bad data, “lack of injuries” or new scouting info.  In fact, that might be a fun article!  I could call it “Ten Things I Think MGL is Wrong About” (and I’d be prepared to have egg on my face at the end of the season).

While I totaly agree with your reasoning and that there is no particular need to change the projection for what is gona happen for the rest of the season i think the title is somewhat misleading.

For teams that are around or even just not to far from “the sweet spot” one would think that the chance of making the post-season would be greatly affected just by the “games in hand”. I two teams were originaly projected to be tied for the division title and nothing special has happened to change the projection and now one team is 13-7 and the other is 7-13 this should make the 13-7 teams clearly favored simply because they have 6 more wins right NOW.

Bjorn, granted that sometimes MGL’s verbosity may obscure his point, he did say:

...simply take their current win/loss record and then “play out” the rest of the season, using the same, damn projections we used before the season started for each team

which is exactly the same thing you said here:

...and now one team is 13-7 and the other is 7-13 this should make the 13-7 teams clearly favored simply because they have 6 more wins right NOW

The basic point is that the future unplayed games should barely change based on what’s happened in the past 20 games, since you’ve got a base on some few hundred games that don’t appear in the 2008 standings, but do appear in the 2005-2007 “player” standings.

Anyway, we’re all agreed.

***

One point though is that as some teams get in the sweet spot, teams will trade their future for a run at it now.  And so, a team could change their roster based on where they are in the standings.  It’s not as if the chance of Yankees giving up their prospects for talent today is the same as the Royals.

But, that’s just a wrinkle.

MGL:
I’m not sure I agree with your hypothetical example.  With 4,000 PAs, the SD for OBP is about .008 (not sure about OPS).  So the 100-point improvement is perhaps a 10 SD improvement (assuming average pitching).  If I saw that, I would guess that something HAD changed to cast doubt on my prior projection.  It might be the overall run environment, rather than a change in player talent, but something has probably changed. 

A comparable sample would be about 4 days of play by all MLB teams.  I wonder how often a collective OPS of .850+ has been posted over 4 consecutive days, after adjusting for pitcher quality and park?  Has it ever happened?

Interesting timing - the writer at Vegas Watch (I linked my name to it) reviewed the last 5 years of pre-season PECOTA team projections and found that in-season performance (as measured by pythW-L) does deserve some weight at this early juncture (just a bit more that a tenth of how much weight the pre-season projections should receive).

I’m inclined to agree with MGL that in most cases, the first 20 games are pretty much meaningless, but there might be exceptions:

1. For players for whom there was limited data before the season, a very hot start probably deserves greater weight. Think Johnny Cueto (should we really go back to his pre-season projection when he has a 29/3 K/BB ratio?). If a team has a few of these players and they’re all trending in the same direction, that might change things going forward.

2. Whatever pre-season projections you make are based on some sort of assumption about playing time. The playing time projection going forward may already be radically different due to injuries or managerial preferences that have become clear in the first few weeks (i.e. dusty baker seems to be sticking with joey votto).

It’s easy enough to figure out how much weight to place.  Figure out how many “career Marcel PA” someone has (as 0.8 times his 2007 PA, 0.8^2 times his 2006 PA, 0.8^3 times his 2005 PA, 0.8^4 times his 2004 PA, and so on...).  Do the same for IP (but use 0.7 all the way through instead of 0.8).

How many “PA” and “IP” do you get?  Let’s say you do it for all the hitters.  You’ll get, I dunno, say 12,000 PA for hitters and 1500 IP for pitchers.  That would be roughly “2 full seasons” for hitters and “1 full season” for pitchers.  So, a Marcel forecast would comprise, roughly of 1.57 seasons (take 57% of hitters and 43% of pitchers).  That means 254 games.

After 25 games in 2008, that means you should weight the forecast 10 times more than the regular season.

All numbers for illustration only.  You can just compute it yourself.  Hopefully, somebody out there will volunteer and do so.

Guy, #5, that is why it is virtually impossible to happen for out hypothetical 1000 players in the lineup team.  But, if it did happen, if we were extremely (like 99.999999999999%) confident that the player performance were independent, we would have to conclude that it was a once in a gazillion random fluc.  It would be like throwing 10000 heads in a row on a coin that we were sure is fair.  Now, in reality, we would be forced to conclude that either the player performance for that one game was not independent, or something was wrong with our original projections, etc., but my point still stands about the difference between 4000 PA from the same batter and from 1000 different batters.  The effect on our projection for this team based on those 4000 extra PA is completely different than if we had an extra 4000 PA from one (or a few) players.

Obviously our projections from this point hence are influenced by:

1) our projection mistakes
2) injuries and playing time
3) whether teams change their personnel for whatever reasons. 
4) changes in true talent.

#6, 20 IP for debut players like Cueto, are STILL not going to change their projections very much, whether you have lots of minor league data on them or not.  If you don’t believe that, just go back in history and look at “the rest of the season” for all debut players who started out great and everyone was talking about, like Cueto.  You are going to find that they are a tad better than any other debut pitcher, and a tad better than any other debut pitcher who started out bad (although obviously the ones who started out bad don’t get much playing time over the rest of the season).

"If you don’t believe that, just go back in history and look at “the rest of the season” for all debut players who started out great and everyone was talking about, like Cueto.”

This is essentially what I did, but in a more objective sense, and for whole teams rather than individual players.  What I found is that 20 games into the season, we should use 90% initial prediction, and 10% Pythag record for each team thus far.  These numbers seem to decrease/increase linearly, by about .25% per game, throughout the season.

If you go back to my post 7, and keep the “250 games” as the fixed weight for the past performance, you get this:

Played Fixed “Played / (Played + Fixed)”
20 250 7%
40 250 14%
60 250 19%
80 250 24%
100 250 29%
120 250 32%
140 250 36%
160 250 39%

Maybe it shouldn’t be 250, but 200:
Played Fixed “Played / (Played + Fixed)”
20 200 9%
40 200 17%
60 200 23%
80 200 29%
100 200 33%
120 200 38%
140 200 41%
160 200 44%

Compare those % numbers to what VegasWatch has:

Played Fixed Tango Vegas
20 200 9% 10%
40 200 17% 14%
60 200 23% 19%
80 200 29% 14%
100 200 33% 31%
120 200 38% 37%
140 200 41%
160 200 44%

Like I said, it’s easy enough to derive whether it should be 200 or 250 or whatever.

Oh, and natch, the more rookies you have, the less that “200” number will be.  You’d REALLY want to do it on a per-team basis.

What I could incorporate into my data is either average team age, or average number of years experience- we could see if a younger team’s first XX games should be weighted more than an older team’s.  Is that kind of data available somewhere for the last five years?

You really should do a progressive weighting, since having 10,000 PA from Frank Thomas doesn’t really mean more than 5,000 PA from Vernon Wells (or whatever those guys have).

But, you have to use it in conjunction with the forecasting system.  Marcel for example, uses 0 PA for minor league data.  With PECOTA, MGL, and Chone, no such luck.  You have to give some weight to their minor league PA, since the forecast is based on that.

In short, you are trying to infer the average regression toward the mean that each forecasting system uses.  It’s not based on age or years of experience per se.

OK, while I certainly agree w/mgl’s discussion on an individual basis, it seems to me that you have to take into consideration the quality of the opposition for each team regarding the 20-game record.  It seems obvious to me that if a team was 7-13 vs only Cle, Bos and NYY that would have a whole different meaning than if they were playing KC, Bal and Tex.  And therefore, I don’t think it’s right as mgl says to just take their 20 game record and “play out” the rest of the season.
And while I’m at it, mgl, why do you think it’s important to add your “if anyone wants to bet, I’ll take all bets at 3-2” to your posts?  A.  Crap, I’ll take either side of almost any reasonable projection at 3-2!  What’s up with that?  B.  If you really want to find out if your projections are so good, I’ll take over or under and -105 on my choice of any 15 teams (avoiding too much cherry-picking).

The 90%/10% weighting is not something I would blindly use.  One, it really depends on the pre-season projections - how good they are and what the criteria used for the projections.  Also, each team is very different.  Obviously if there are major changes in the presumptions that went into the projection, then the 90/10 “rule” has to be altered.

In any case, 90% pre-season is a lot!  And even then, we are talking about the other 10% being pythag and not actual record.

If I am doing the final projections, I don’t see any reason to care about the actual or pythag record so far, at ANY point in the season, any more than I care about how a player has done recently in doing a player projection, other than how that recent activity changes the projection.

Same with team projections.  If I am doing a team projection, I simply project the players (using current season performance to update the player projections of course) and their playing time, and “play out” the rest of the schedule.  I don’t see any reason to pay any attention to the regular or pythag record so far, at any point in the season.

Now, for a forecasting system like Pecota, which does not project things like baserunning, and defense very well, at least not in the past, then you MIGHT have to use a little bit of season “team” performance to tweak the rest of season projections, but I still wouldn’t use something as gross as actual or pythag w/l records.  I would use things much more granular, like team offensive, defensive, and baserunning lwts, and pitcher ERC’s, FIP’s, or whatever.

I am amused to note that Johnny Cueto’s performance is now below, for example, PECOTA’s projection for him. Does anyone want to suggest that we should now revise his projection substantially downward? Probably not. Small samples are always small samples, absent extrastatistical evidence. Period.

#16, well said!  If people kept track of what everyone (the media and talking heads mostly, or as Al Michaels says, “gasbags on parade") says about players and teams during the season and then looked at those at the end of the season, it would be hilarious.  All the players who have “turned the corner” or were “done” or the teams that are “a lot better or worse than everyone thought” (although I am REALLY tiring of commentators saying, “NO ONE thought that the Rays would be this good.” Yeah, NO ONE from the media - look at the analysts projections you idiots!).

I do have a good projection for Cueto though (his MLE K and BB rates were very good in Chatanooga last year), and as of a week or so ago, he had pitched to a 3.85 NERC, which is very good for a rookie starter (major league starters are around a 4.00).