ericvman and jmei debate player projection

Haven't done the individual player YTY correlations. That's the next step. However, that's just one of the things you're looking at.

Standard deviation / mean, annual MLB rates, 1980-2013:

.016 1B%
.025 BABIP
.065 XBH/HIP (aka XHP)
.069 HRX
.108 XBC
.166 HRC

There has been a lot more YTY variation in overall MLB HRC rates than XBC rates. So when we start with HRC, we are overlooking the evidence that XBC is more fundamental, and that HRC is basically double-counting hardness of contact (HOC) because as XBC rises, so also does the proportion of XBH that leave the park.

Inter-correlations (across the 34 years)

HRC and BABIP .873
HRC and XHP .913
XHP and BABIP .861

For some purposes, this is good. The three metrics really tend to move together from year to year. But what I believe you're doing here is taking generic HOC, and parceling it out into three metrics, each of which also contains some specific skills as well.

HRX and 1B% .533
XBC and 1B% .631
XBC and HRX .866

This looks like it does a much better job of isolating three different skills, and of course logically it does, too. Jose Iglesias really didn't have a fluke BABIP year; he had a fluke 1B% year, and future efforts to measure his actual talent at infield hits will work better looking at 1B% than BABIP.

What is really interesting is that in predicting the annual MLB K rate, they yield very different regression formulas, although both do an excellent job. The prediction using HRC and BABIP has no terms for annual BB rate; its influence on K rate is no longer discernible when you're using three different measures of general HOC. The regression with XBC and 1B% does include BB rate, and uses fewer HOC variables (including interactions).

But what really grabs me is this: in neither regression are terms for "steroid era" significant. Yet the regression using HRC and BABIP is generally more accurate, while the regression using XBC and 1B% kicks its butt in two places: the steroid era and the run-up to the 1987 HR peak. I'm looking into that further today, and then I hope to write the K rate prediction stuff up for publication.

First, a point: HR are even more influenced by park than 2B and 3B are. I think the number of balls that are ambiguously singles or doubles, depending on the player's speed, is actually fairly small; certainly the influence of speed on doubles is less than the influence of swing trajectory (FB/GB rates) on HR, given equal hardness of contact. So the exogenous factors on HR are as large or larger.

Now, I'm really good with Excel ... it took me 23 minutes to get my first set of Y2Y correlations and 2.5 hours to do about as thorough a job as I can imagine.

Here are the correlations for every possible contact metric, 1980-2013, minimum 400 PA (if you'd like to know some other time frame or minimum number of PA, that would be trivial to do). I've bolded the conventional metrics and put my proposed alternatives in red.

.870 K%
.804 HRC
.803 HR/Hit
.777 BB%
.759 HR/XBH
.739 XBH/C (XBC)
.721 XBH/H
.525 CBA
.484 1B/C
.451 XBH/BIP
.442 BABIP
.426 XBH/HIP
.420 1B/(1B+OIP) aka 1B%
.419 23/C

HRC does have the best YTY correlation, but it's the product of two things that I'm proposing should be separated, both of which have correlations 92% - 94% as strong. (The product of two things which correlate with each other, each of which has a high correlation across time, I think, always yields an even higher correlation.) If you take HRC, you are left with one of two low-correlation pairs (the other being XBH/BIP and 1B%). My proposal is the only way to get two high-correlation metrics.

OK, let's repeat this for 250 PA minimum. And I'll just report the percentage of the correlation strength retained, which is to say, this is the rank by sensitivity to sample size (least sensitive to most), confounded by whatever extra variance you get by adding less talented, low-PA players to the sample:

.988 HRC
.983 K%
.983 HR/Hit
.981 XBH/C (XBC)
.970 XBH/H
.967 BB%
.960 HR/XBH
.944 CBA
.943 XBH/BIP
.926 23/C
.934 XBH/HIP
.899 BABIP
.893 1B/C
.888 1B%

I've always felt that that 1B% was better than BABIP at isolating luck, and this backs that up. However, overall the new breakdown stands up to small sample sizes better, because HR/XBH doesn't deteriorate quite as much as XBH/HIP. And XBC (which stands up 99.3% as well as HRC does) stands up better than BB%, which is impressive.

And now I'm going to do just the six metrics we're focusing on (plus K% and BB%), minimum 250 PA to get a larger sample size, for five different eras:

1980 - 1987, increasing K and hardness of contact
1988 - 1992, retrenchment*
1993 - 1996, the rise resumes
1997 - 2005, steroid era; relationship of MLB K rate to HOC much less predictable
2008 - 2013, rising K rate (pitch/fx and increased velocity)

*I'm beginning to think that the famous 1987 spike in homers was largely the culmination of a trend of increasing strength from weight training, plus a bit of warm weather, and that MLB deadened the ball afterwards. In this era, K rates (a function of hardness of swing) were higher than predicted from HOC measures and walk rates, indicating that the reduced HOC was being generated by harder swings than the model thinks it was.

Here's a table with the results (BBP = BABIP)

Year N K% BB% HRC BBP XHP XBC HRX 1B%
1980 1130 .836 .760 .787 .398 .341 .684 .733 .394
1988 643 .860 .739 .806 .381 .324 .693 .758 .348
1993 446 .825 .754 .799 .320 .349 .696 .739 .322
1997 1254 .833 .768 .806 .327 .380 .733 .719 .332
2008 808 .851 .694 .731 .328 .373 .658 .657 .360
So what here might not be random?

-- A steady decline in the Y2Y correlation of 1B% from 1980 to the mid-90's. Which is to say a reduction in the skill component (bunts, fast players who get a lot of infield hits). It may have risen again post-steroids.
-- A spike in the Y2Y consistency of XBC (or XBH / Hits in play) in the steroid era, maybe. That would happen if you had a greater stratification of talent, i.e., guys using and guys not using. There was no change in the standard deviation, though. So did PED use somehow reduce Y2Y variation for users?
-- A collapse in the Y2Y consistency of walk rates and hardness of contact (XBC and HRX, or their product HRC) in the exploding-K rate era. I'm not sure anyone has noticed that, and it's very interesting.

I think the argument here is extremely solid. Both XBH/Contact and HR/XBH correlate really well from year to year. They correlate with one another, as well (.678). The extra strength of the Y2Y correlation of HOC doesn't mean it's a more fundamental stat than XBC; it happens because you're adding some flyball / groundball tendencies to XBC. If you put all your eggs into the HRC basket, you're left with two noisy stats. One of those, BABIP, is clearly not as good for isolating the luckiest outcomes as 1B% is, conceptually or empirically.

In short, what hitters try to do is hit the ball hard. Yes, there are lucky doubles and triples, but most are hit significantly harder than singles. If you excel at hitting the ball hard, you're going to have more XBH hits per contact, and a higher percentage of those XBH are going to be home runs. It's answering two questions: what percentage of the balls you hit were hit hard? What percentage of the hard hit balls were hit really hard (and up in the air)? That gives you a double scale. You can differentiate guys who consistently hit the ball hard, but not super-hard (e.g., Pedroia) from guys who crush the ball. (Adding FB/GB data would do an even better job, but for the time being I'm trying to limit myself to data that will be available across all historical eras, and has no subjective element*.) Multiplying the two together to get HRC just combines your two more nuanced measures, so of course it correlates even better, but the price that you pay is that you don't have a second non-noisy stat.

*Hard hit balls that first land on the mid to outer dirt part of the infield are often scored as line drives by BIS and ground balls by MLB, or vice versa.

Aha, here, I think, is the smoking gun that doubles and triples can and should be first lumped together with home runs, not with singles.

The within-year correlation of HR/Contact and 1B/Contact (all 6596 players with 400+ PA) is  -.440. Guys who hit more of one tend to hit less of the other. So, no, as we all agree, you want to separate those.

The correlation of (2B+3B)/Contact to HR/Contact is .260. The correlation of (2B+3B)/Contact to 1B/Contact is -.184.  Clearly, they're a form of power hitting, not a form of non-home-run hitting.

No intention of joining this debate, but I find it interesting, Some of it is beyond me, but it's thought provoking. Two smart posters. Carry on and thank you.

Fair enough, but I still think the majority of that importance is captured by regressing HRC alone to career norms. The fact that a player's XBC stayed the same as well doesn't tell you that much more, I think, at least in the age range that guys like Saltalamacchia and Napoli are in (i.e. age-related decline shouldn't be catastrophic yet).



I think this is a spot-on distinction and gets to my skepticism of 23C. I agree that there is some element of 23C that gets you to the "hits the ball hard" stat that you want to find, but there's enough noise that I'd rather just rely on some variation on home run rates. This is where even limited HITf/x data would be supremely useful.