Hi, I have a basic question that I just couldn't find an answer for.
I want to measure the % error introduced by using EWMA as against a linear average for a *stationary* random process (not necessarily Normal) over a given (long/short term) time window: I am using the Chi Squared test for anomaly detection on a computer network. In the setup, sampled values of a certain network parameter are distributed into one of 32 mutually exclusive and exhautive bins. The long and short term frequencies of each bin are then calculated and used with the Chi Squared test. The frequency calculation involves the use of EWMA since other methods are not computationally feasible. I wish to know how much of the False Positives that we are getting is due to the inaccurate estimation by the EWMA of these frequencies. Surely, the process is not Normal, there *is* serial autocorrelation and since it's a Quality of Service Network Domain, the process can be assumed to be fairly/approx.ly stationary. I can see with my code and results that the EWMA tends to be near the actual 'brute force' linear average, but not quite equals it. Since this difference matters in Chi Square, it will lead to False Positives. I just want to know whether the error is significant at all. Can anyone point me to a detailed mathematical treatment/paper on this topic? I couldn't find any. Thanks a lot! Vinay. -- Vinay A. Mahadik http://hickory.csc.ncsu.edu ================================================================= Instructions for joining and leaving this list and remarks about the problem of INAPPROPRIATE MESSAGES are available at http://jse.stat.ncsu.edu/ =================================================================
