another thing you could try is using "fminsearch",
please have a look at the example of Matlab help
on this function.

(sorry typing English is easier for me)

basically you have to think about what kind of function
(or series) you want to approximate the "real data" with.
then it's not too hard to find the "best" parameters(usually
coefficients) in the function/series that can represent the real
data with "least" error.

I just helped my friend solving this kind of problem, the
code required to do this isn't too hard to write.