THE BOOK--Playing The Percentages In Baseball

You know, it makes more sense for #1 to be ERA and #2 to be W/L.  It was Verlander/Weaver all the way (i.e., low ERA) and the tie-breaker was W/L.

So, I amended the title.

That’s some pretty tortured logic.
For the top four, they voted using ERA, with wins/Ks as the tie-breaker:

V: 2.40 (24W)
W: 2.41 (18W)
S: 2.82
C: 3.00

Sorry - my post was in progress are you posted your amendment.

It nicely converges with the NL findings and makes some sense: voters opt for some medium between true talent (FIP, Ks) and ultimate team- and luck-facilitated outcome (Wins)

The ERA difference between Shields and CC is very close.  Otherwise, you’d have to include a bunch of other guys between Shields and CC.  CC did not finish #4 in ERA.

It is almost like the voters weigh 1 win as 5 pts of ERA. This seems to work for the order of the starters in the voting, with the exception of Haren beating Romero, where is one more win was worth 25 pts of ERA. Perhaps this is the small market bias coming through. Not sure.

By the way, this is how the Cy Young predictor gets created.

It’s fairly easy to go through this, each year, each league, and come up with a points system.

For all the hemming-and-hawing of BBWAA voting like human beings, we can distill their voting patterns based on how they interpret numbers.

I’d like to see an aspiring saberist come up with a new Cy evaluator based on the past 10-15 years.

You can throw in parameters like “heart” or “saves was a big story this year”, or whatever.

The problem with a Cy predictor, as has probably been noted, is that standards can change over time. So, you’ll never have enough sample size to establish a model that is at all stable. My guess is that wins is less important that it once was (e.g., Felix vs. Welch), for example.
It’s too much of a moving target…

The Cy Young predictor almost nailed the top 5; the only difference was that the formula predicted CC 3rd, Shields 4th, but Shields beat CC to 3rd 66-63.

So I would suspect they’re still voting the way the formula predicts.

The formula is:

((5*IP/9)-ER) + (SO/12) + (SV*2.5) + Shutouts + ((W*6)-(L*2))

So, since all starters got about the same number of decisions (26-29), each win is worth about 8 points according to the formula. The difference between best win/loss and worst win/loss among the top 4 guys is 50 points. (134 for Verlander, 84 points for Shields.)

If each pitcher has around 225 innings pitched, then a 0.1 difference in ERA is worth about 2.5 points. The difference between top and bottom in ERA (2.4 for Verlander vs. 3.0 for CC) is ~15 points.

The difference between top and bottom in strikeouts is 4.3 points.

The difference between top and bottom in shutouts is 2 points.

And innings pitched, since most Cy Young contenders have ERAs around 2.75 +/ .25, is worth about 0.25 points per inning. The difference betwen top and bottom here is 16 IP, or about 4 points.

So it would normally be #1 W/L (by a mile), #2 ERA, #3 IP or K. But Shields edging out Sabathia may mean they’re now putting greater weight on ERA, IP or shutouts than they have in the past.

But if past Cy Youngs have overweighted W/L as dramatically as the formula says, then I’d be skeptical of any claim that ERA is now #1.

I checked to see if I could find a formula that would retroactively select the recent winners, and it’s impossible.

It’s impossible for Bartolo Colon to beat Johan Santana in 2005 AND for Felix Hernandez to beat CC Sabathia in 2010 with a single linear formula (without some sort of “heart” fudge factor).

Bartolo Colon is ahead of Santana by 5 wins, but behind by 9 IP, 81 K, 2 shutouts and .61 ERA.

Hernandez is behind CC by 8 wins, but ahead by 12 IP, 35 K, .91 ERA and 1 shutout.

.61*8/5 = .976, 9*8/5 = 14.4, 81*8/5 = 129.6

So if Colon is ahead of Santana, Sabathia should be ahead of someone who has Hernandez’ W/L record and is up to 14.4 IP, 129.6K, 3.2 shutouts and .976 ERA better than him.

I guess you’ll have to use a nonlinear predictor. But then it’s guaranteed to overfit.

With only 20 datapoints, it was going to overfit anyway.