I suggested a few weeks ago that we try to come up with a way to measure the accuracy of the Streetcar NextBus predictions (since they have been a source of some complaints reported in the press).
Turns out someone has already done it (at least for a small data sample). Rob Bertini of PSU forwarded this student project (PDF, 307K) by Hiu Ng from last year.
As I read his data, the most common result seems to be that Streetcar arrives about 1 minute after the predicted time. If you think about how you would want to bias prediction error in such a system, if definitely seems like you’re better off having people arrive a little early than risk having the car already departed by the time you tell them it’s going to arrive!
So I think this confirms my anecdotal experience, but I’d still love to find a way to collect a larger data set.