I'm estimating a deadline, when my time series will add up (total so far) to a certain large number. I'm doing so by getting a forecast line, plus or minus the RMS error of the known values. But if I the error every period, and subtract it every period, to get the upper and lower bounds on my estimate, that will yield a worse estimate of my deadline than is likely. How do I find the error more reasonably? Should I instead use the RMS error itself, not compounded many times, since over time the underestimates and overestimates will tend to add up to something within the RMS error? (Something like that reasoning, anyway. The series is generally decreasing, so maybe that's a reasonable idea.) Any better ideas?
You might want to pursue http://autobox.com/cms/index.php/blog/entry/will-we-make-the-month-qtargetq-number as it deals with your kind of question. If you post your daily data in a csv file and define what the data is and define your target # I will try and make sense of it.
I answered a similar question ( at least similar to me !) here Confidence Interval derivation and received a down vote without any explanation or reflection. It is, at least to me, kind of counter-productive not to explain why the down vote thus consequently not providing positive feedback or a viable way forward.
You might consider re-titling your question in the spirit of my initial comment as your problem should have more of a spotlight and get more attention.