Ah, a statistics question!

Best is relative:

1. Is the model that got 5 storms perfect but missed the remaining 20 storms by over 500nm to the south the best?
2. Is the model that never got one right but got half right by being off exactly 300nm north and the other half exactly 300nm south best?

The first one you had 5 right, but the average is skewed way south so you're likely to get bad predictions southward with the storm, thus it is inaccurate.

The 2nd one you know you are always in the envelope, and by manipulating statistics, you can even say that, "the model predicted the average path of every storm correctly," becuase (300N + 300S)/2 = (0N)/2 = 0nm off.

---

Now, on to the opinions

GFDL, when it was correct, was very correct. When it was wrong, it was very wrong. It was more often correct than wrong. GFS and NOGAPS were ok, not great, but not bad. BAM was often really bad, but occasionally correct (usually when GFDL was wrong). CMC was erratic - sometimes right sometimes wrong and no commonality when.

That's my observations. Anyone have anything official?