Friday, May 18, 2007

Why have the Reds deviated from Pythagoras?

One of the puzzling things about the Reds' record this season is that they're well below where they should be according to their Pythagorean Record, which estimates where the record should be based on a team's runs scored and runs allowed. Here's the breakdown (thanks to The Hardball Times):

Actual record: 16-25
Runs scored: 181 (4.4 r/g)
Runs allowed: 189 (4.6 r/g)
Pythagorean record: 20-21

In Dave Studeman's column today, he uses a simple method that he developed last season to assign the blame (and credit) for deviations from a team's Pythagorean record based on win probability data. Basically, he uses a regression to estimate what a teams' batting (or pitching) win probability should be given how many runs it has scored (or allowed), and then compares that to its actual win probability. If a team scores a lot of runs when it doesn't count, but does not score runs when it does count, they will show up in this analysis as underperforming.

Studeman's analysis put the majority of the blame (2.4 wins below expected) on the Reds' offense. Pitching was tabbed with "only" 1.0 wins below expected, given the number of runs they have allowed. This goes a bit against conventional wisdom in the Reds' blogosphere right now, which seems to assign most of the blame to the Reds' bullpen. Of course, if the Reds' starters have had a more positive influence than is typical for their runs allowed, while the bullpen has had a more negative influence than is typical for their runs allowed, the effect of the bullpen might be hidden in his analysis.

To tease the effects of starters and relievers apart, I replicated Studeman's procedure, regressing batting WPA on runs scored, starter WPA on runs allowed by starters, and reliever WPA on runs allowed by relievers. All WPA and runs data came from fangraphs.com, and were current through May 16 2007. Here's how it breaks down:
Teams Batter Deviation Starter Deviation Reliever Deviation Predicted Pythagorean Deviation
Actual
Pythagorean Deviation
ATL 1.7 -0.9 2.4 3 3
DET 0.9 0.8 1.3 3 3
CHA 3.1 -1.2 0.6 3 3
TBA 1.8 1.0 -0.4 2 3
MIL 0.2 0.4 1.6 2 2
CLE 2.2 0.4 -0.7 2 2
SEA 2.2 -0.4 0.0 2 2
ARI 1.0 0.4 0.3 2 2
LAN -0.3 -0.4 2.4 2 1
STL 1.8 -1.4 1.1 1 2
COL 1.4 0.6 -0.9 1 2
PIT -0.7 -0.6 2.2 1 2
LAA -0.3 1.2 -0.1 1 0
NYN 0.5 0.2 -0.2 1 0
WAS -1.4 -0.5 1.5 0 1
BOS 1.2 -0.6 -0.1 0 0
HOU 0.3 -0.6 0.3 0 0
FLO -0.8 -0.8 0.7 -1 -1
TOR 0.0 1.9 -2.7 -1 -1
BAL -1.5 0.8 -0.6 -1 -1
MIN -1.7 -0.6 0.9 -1 -1
SDN -2.6 0.5 1.3 -1 -2
PHI 0.8 0.3 -2.6 -1 -2
TEX -1.3 -1.3 0.9 -2 -1
KCA -1.7 0.5 -0.6 -2 -1
SFN -0.4 0.6 -2.1 -2 -2
OAK -1.5 0.7 -1.5 -2 -3
CHN -2.0 0.3 -1.6 -3 -4
NYA -0.3 -2.2 -1.3 -4 -4
CIN -2.6 0.8 -2.0 -4 -4
The deviations noted above are essentially the differences in expected wins attributable to hitters, starters, and relievers, given the number of runs they've scored or allowed. As you can see, there is good agreement between the predicted deviation in wins from my regressions and the calculated Pythagorean record. There are inconsistencies, but they are never by more than a win...and keep in mind, we're dealing with only ~41 games here, so we'd expect a few more misses we might later in the season due to the small sample size. Also, if you compare my hitter deviations to Studes', they match up well--I think the small differences you see are probably just due to the fact that I have an extra day's worth of data in my dataset.

This analysis shows that the Reds' offense and bullpen are both to blame for the team's under-performance of its Pythagorean record. In fact, the offense is tied with SDN for the largest negative residual in WPA of any team in the majors right now, while the bullpen ranks 4th from the bottom.

There is a 2.8 win difference between the performance of Reds starters and relievers, relative to where their overall rate of allowing runs predicts they should be. This is not uncommon, indicating that it is useful to consider starters and relievers separately. Other teams with equal or larger differences between starter and bullpen effects on Pythagorean Record include Toronto (4.6 wins), Atlanta (3.3 win difference), Philadelphia (2.9 win difference), Pittsburgh (2.8 win difference), and the LA Dodgers (2.8 win difference).

Here are the data graphically. The deviations listed above refer to distance above or below the regression line (i.e. the residual):

Hitters (R2 = 0.66)
Starters (R2 = 0.73)Relievers (R2 = 0.35)The good news is that this probably means that the Reds can be better than their record indicates over the rest of the season, as deviations from season-average performance in high-leverage situations don't tend to be consistent over time (i.e. "clutch" players in the past don't tend to be clutch players in the future). The bad news is that bad performance in the most important situations by the offense and the bullpen already has the Reds down 4 wins below where they should be, which makes a playoff run a fairly low probability venture at this point in the season.

2 comments:

  1. Most teams aren't going to hit it right on. The theory is not perfect. There is a luck factor but one thing not stated is that will eventually be made up. If a team finished 74-88 and should have been 78-84 it will be made up in the next season or two. There is a consequence for everything--universal law but the consequence doesn't always happen righ away. In 2006 Cleveland should have been 89-73 and they were something like 11 games below where they should have been. In was a cinch that they were going to win the AL Central in 2007. Tampa Bay was well above what they should have been so don't expect a lot in 2009. The Angels were 12 games over their expected and they will eventually have to pay for that. The Tigers and Cleveland consistently finish below their expecteds but eventually will overachieve to make up for the bad luck they've had. Minnesota has lately had good luck and their division win in 2006 was mainly luck--they should have finished three games below the Tigers instead of a game above. Minnesota was swept in the first round and it appears they are having to pay for it now--little by little.

    ReplyDelete
  2. @Anonymous,

    I don't disagree with anything that you're saying. I think we're just answering the question on two different levels. To compare this to some biological jargon, you're answering it on the "ultimate" level and I'm answering it on the "proximate" or "mechanistic" level.

    Essentially, this article was about clutch--how well does a team's raw numbers compare to their WPA numbers. That's not to say that there is repeatable signal behind these deviations--it just lets us know specifically where deviations have occurred. That's what my last paragraph was about.
    -j

    ReplyDelete