I am trying to figure out the optimal combination for the size of data set (n), number of prediction coefficients (m) and the size of predicted data set(f). Is there any logical reasoning or practical considerations? I've searched the literature but the only recommendation I could find relates to predictive coding (m<<n) with LP based on covariance. I need f>= 64 for future set but large "n" and "m" are not realistic for real-time DSP.

I am sure that this question has been analyzed but I cannot find the answer. I'd appreciate any help.