Or rather there was – it was fixed in a later version and I simply didn’t publish those yet. I’m still revising the Marcels code that I have – all rate stats should be unaffected. I’ve been putting off publishing revisions until the Baseball Databank gets released, but since these are getting as much play as they are I’ll see about publishing the revised version later this afternoon.

Here’s how Marcels regression works:

x/(x+214)

Where x is number of PAs in sample (in this case weighted PAs). Instead of 5/4/3 weights I use equivelent fractions, like 1/.8/.6. Keeps things a lot tidier for me. (It all works out the same in the wash.)

So for a player with 300 PAs in sample, you get:

300/(300+214) = .58

Which means that he gets regressed to the mean 42% (1-.58). For a player with 1600 in-sample PAs (a full-time starter, in other words), you get:

1600/(1600+214) = .88

So only 22% regression to the mean.

That is, in fact, reflected the in “R” column in the spreadsheet – that lets you know exactly how far each player’s stats were regressed. Bonds missed a full season, and the one that’s weighted the most heavily, so he gets regressed more than someone like, say, Pujols.

Marcels more drastically regresses playing time, and it only uses two years of data, like so:

.5 * PA1 + .1* PA2 + 200

Where PA1 is PAs in 2008 and PA2 is PAs in 2007. That really lowers the playing time forecast for someone who didn’t play at all in 2008.

[As a final aside – any forecasting system is likely to do better with projecting rates of performance than projecting playing time. The optimal approach involves combining the rate stats from a projection system with a better projection of playing time based upon depth charts and such.]