I think it’s important to note (as may have been done; I wasn’t around when the projections first came out) that especially in Marcel’s case, the projections for the young (experience-wise) guys are barely worth noting. Gaby, for instance, has a reliability score of .03 — in essence, the projection is 3% him and 97% regression. Which is to say that, lacking MLB data, the assumption is that he will be an average 26 year old MLB player. With CHONE, you won’t see it in the weighted average, but if you dig into the individual pages, you’ll note a much, much wider spread of numbers between the 10th and 90th percentile projections, reflecting the lack of certainty. PECOTA is similar, though with a better “mean” to regress towards, based on its physical comparisons.

The system that should be intriguing for the Marlins is the still relatively new Oliver system by Brian Cartwright at Statistically Speaking. Oliver is essentially Marcel, but taking all seasons, major and minor, into account. Now, the inclusion of minor league data makes a big difference for young guys, but it also relies on MLEs. Obviously MLEs are less than perfect, but I think especially as time has gone on, they have become about as accurate as can be. Of course, this cuts both ways, so while Oliver will generally be more positive on young starter-worthy players than Marcel, it will also look less favorably on someone like Hanley, whose boredom with minor league ball is both well document and apparent in his stats.

The basic lesson is this: Marcel is intended to be the absolute least “intelligent” projection possible (that is, the one that takes the least amount of things into account) and as such, will strongly push inexperienced guys towards average; and similarly, all projection systems are built on top of regression, and as such, dampen the spread of numbers. Essentially, because it’s both futile and unrecommended to project a specific player to put up a specific line, the idea is either “this guy has X% probability of doing this” (with PECOTA and CHONE) or “of the group of guys projected to hit X, about half will do a bit better [with perhaps one or two exceptionally better] and about half will do a bit worse [with perhaps one or two exceptionally worse, likely due to injury]” (with Marcel and Oliver). They are projections, not predictions, and as such, try to mark the center for what will naturally be a spread of possibilities. That’s why you’ll see that, for example, PECOTA projects only 11 players to hit 30+ HRs. Is that because the people at BP only think 11 guys will hit 30+ HRs? No, it’s because, taking all risks and possibilities into account, there are 11 guys who are most likely to hit 30+. That’s the difference between a projection system and somebody like us making a guess at what someone will hit: we don’t say “well there’s a 10% chance he gets hurt for the year, so let’s knock him from 40 to 36.” That may work with that one guy, but naturally, it’ll be way way off on 10% of guys (I’m just making up the idea of a 10% attrition number of course). The goal of the systems is to be accurate across the spread of players, rather than nailing any one guy.