Only seven players ever got more than 100 PA as a teenager. So these are Mike Trout’s only comparables, along with the number of PAs they got in their second season after reaching the major leagues:

Adrian Beltre: 614
Andruw Jones: 467
Edgar Renteria: 691
Alex Rodriguez: 149
Ivan Rodriguez: 123
BJ Upton: 0
Justin Upton: 417

The average is 351 PA, and that’s probably why PECOTA projects a “half season” for him. Of course the variation is huge; Renteria and Beltre played every day, while BJ Upton spent his entire age-20 season in the minors. But this tends to happen with a sample size of seven. That’s just the nature of a projection though, it’s not a “prediction,” it’s a statistical average with a varying degree of confidence, depending on how much data is available.

Yes, it would be nice if we had a more sophisticated system. Think about what you’re proposing though. If you start making forecasts for one player based on other players, then where do you stop? Is it really that obvious that Mike Trout will get whatever PAs Vernon Wells doesn’t get? You can’t factor in Mike Scioscia’s love for gritty veteran-ness, because then you have to make similar judgments for every other manager. It’s also possible that Vernon Wells has a pretty good season; that has to be accounted for too. Maybe he only gets used against left-handed pitchers, which would inflate his batting numbers. How would that reflect on Trout’s projection?

Suppose Mike Trout gets injured, or suppose DiPoto decides that he’d rather trade for an experienced outfield mid-season rather than push the rookie. Then it gets even worse. Now you have to make decisions like this for every player in the league. Mike Trout’s playing time doesn’t just depend on Vernon Wells, it depends on every other player who might possibly fill that position, and all of their projections also depend on Mike Trout. This just isn’t a tractable problem. Even if you could do it, you’d probably get a number not very far off from the PECOTA average, because of all the uncertainties.

I’m pretty sure this is why the people who make things like PECOTA just select a method and stick with it for everything. Otherwise you end up having to justify millions and millions of arbitrary assumptions about what counts and what doesn’t. These are all proprietary systems, though, so I admit that I don’t really know what’s going on in detail. This is just based on my own experience with other kinds of modeling. Dan Szymborski seems to have radar for this kind of discussion, so maybe he’ll chime in with his thoughts.

I just don’t feel that it’s a great use of resources to do that.

Computer projections are the most useful when they’re doing what computers do best: finding patterns in macro data and putting what the numbers mean into context objectively. Humans are wired rather poorly for this task – we’re biologically designed to see patterns where none exist.

For the situation on the ground at the micro level, a human can better deal with the more intangible aspects of playing time. My policy for projections is to let computers do what computers do best and let humans do what humans do best.

In fact, I feel that having the computer pick actual major league time is not only a futile task, but possibly counterproductive as it removes information from the equation. I want a computer to give me a data-driven opinion about how good Alexi Amarista or Tyson Auer are, not that Amarista will get a little bit of playing time and Auer none.