Implementing Moving Average Algorithm for Data Analysis

[khdani]

hello,
i'm trying to implement the moving average algorithm
http://en.wikipedia.org/wiki/Moving_average

i don't understand, after i calculate an average of say 5 values, what do i do
with the x values? how do i plot the new data points?

 

[chroot]

The output at time n is equal to the average of the input at times n, n-1, n-2, n-3, and n-4.

- Warren

 

[khdani]

thanks chroot,
is there a better algorithm, which gives more accurate results ?

 

[minger]

Define accurate. What are you trying to represent?

 

[khdani]

i've spectra graphs of this kind:
http://img91.imageshack.us/img91/692/60co.jpg

i've smoothed it to this:
http://img37.imageshack.us/img37/1558/smoothed.jpg

but i want it to be without this small fluctuations, as linear as possible, but only the noticeable peaks to keep.
so, is there any other techniques except moving average ?

 

[minger]

Increase the "sample size" of your moving average. That should smooth it out some more.

 

[skeptic2]

Look carefully at the peaks of the two graphs. The one with the running average has the peaks shifted to the right. The larger the sample size, the more they will shift to the right. You might consider moving your whole graph to the left by half the number of points in your running average.

Or consider averaging a point with the same number of points to the left and to the right of it. This method has the disadvantage of averaging fewer points at the left and right extremes of your data.

 

[skeptic2]

but i want it to be without this small fluctuations, as linear as possible, but only the noticeable peaks to keep.
so, is there any other techniques except moving average ?

Do you know to implement a high pass filter?

Are you working in Excel or programming this in a particular language?

 

[khdani]

i'm programming this in C#,
i don't know how to implement high pass filter, would you please explain what exactly do you mean by implementing a high pass filter and would it help?

Or consider averaging a point with the same number of points to the left and to the right of it. This method has the disadvantage of averaging fewer points at the left and right extremes of your data.

i didn't understand this, would you please explain this ?

 

[minger]

That will depend on whether you're doing this in real-time. If you are, then you can only use data that you've already taken, or in the past. That will force you to use a backwards method. If you are doing this post-processing, then you can average from both behind and in front of your point of interest. This will keep the peaks at the right place in time.

Alternately you can look at CFD dissipation algorithms which are similar in a way to what you're doing.

 

[chroot]

Well, you'd want a low-pass filter, not a high-pass.

A low-pass filter is a system which permits low low frequencies to pass from input to output, but blocks high frequencies. The small wiggles on your plot are high-frequency information, and would be blocked by such a filter.

If you don't know how to design one, it's really too complex a topic for us to explain here. If you have access to MATLAB, though, you can design some basic filters with just a couple of lines of code, and we can help you with that.

- Warren

 

[skeptic2]

i'm programming this in C#,
i don't know how to implement high pass filter, would you please explain what exactly do you mean by implementing a high pass filter and would it help?

i didn't understand this, would you please explain this ?

Only you know what want and what you consider accurate. If what you want is to locate the peaks without regard to actual values this may be of help.

Again given the raw data points at A(n) and output at B(n), B(n) = A(n+2) - A(n-2) + A(n)*Exp(t).
You may be able to enhance the peaks by changing the value of t or raising the result to an odd power greater than one. Try starting with a value for t of -1.

B(n) = (A(n+2) - A(n-2) + A(n)*Exp(t))^3
Running averages typically average the current value with historical values resulting in a shift of the curve to the right by half the number of data points that are being averaged. To prevent that, the running average must average the same number of future points as it averages historical points. Obviously this isn't possible if you're doing real time processing or averaging at either end of the data series.