As promised earlier today, I went through and collected all the arbitration salary data I could find for this winter. Here's the spreadsheet, with my WAR projections, the projected salaries, and the actual salaries (or the figures submitted by player and team, for those who haven't yet agreed to terms): https://spreadsheets.google.com/pub?key=0AjCZOpYYLVwqdEVmUjlmdllXUWZZWEJzdjE5R2JibVE&hl=en&gid=0. That sheet should include everyone who is arbitration-eligible and projects above replacement level... I couldn't find the salary data for about 20% of the guys on the sheet, but that's okay, adding them in later won't meaningfully alter the results.
Basically, this winter's data jives quite nicely with what I found two years ago. Back then, I came up with a formula which estimated the proper salary progression as 27/41/65% of free agent value for players in their first, second, and final years of arbitration. And as you can see from the spreadsheet, the values for this winter are 27/39/64, virtually the same. (Yeah, I'm lumping super-2s and normal first-year arb-eligible guys into the same category for these purposes... you can play around with the numbers yourself if you want, but trust me, it makes very little difference). This gets us to 99% accuracy or better for each service time class, taken as a whole.
One thing that absolutely needs to be emphasized, though: this formula is not meant to be a really accurate predictor of how much money a particular player will make in arbitration. The reason for that is simple: arbitration awards aren't based on WAR projections, they're based on past performance. And not only that, they're based on a horribly inaccurate view of past performance, which overvalues HR, RBI, and saves while undervaluing defense (this is why Prince Fielder gets $15.5 million and Rickie Weeks will end up around $6 million, even though Fielder only projects half a win better than Weeks). What this formula is designed to do is to provide a general guideline for how WAR translates into arbitration dollars, to provide something accurate in place of the standard but unproven 40/60/80. In other words, this is a rule of thumb, nothing more. But it's a pretty good rule of thumb, I hope.