I'm new here so I will try to be as clear as possible. I am trying to apply some techniques from signal processing framework to denoise financial time series.

I would like to know if what I am trying to do makes sense or if I am committing some theoretical mistake;

I have a time series that I assume to be composed by two sources (non observable), one a genuine signal while the other is just white noise. $$Y(t)=X(t)+e(t)$$

I also assume that the white noise is additive and is uncorrelated with the signal. From here what I am planning to do is to apply the STFT to the observable data.

For every frequency I construct a Hankel matrix (a sort of trajectory matrix for the Fourier coefficients). Then I obtain the variance-covariance matrix by multiplying the Hankel matrix by its hermitanian.

Here is where things become confusing: I want to make the eigenvector decomposition of this matrix, such that I'll get something like:

$$U(\Lambda+\sigma^2I)U'\,\text,$$ assuming that the error term is white noise, and there is no overlap between windows, and $\Lambda$ being the diagonal matrix containing the eigenvalues of $X$ on its diagonal.

If I further assume that the signal $X$ has a smaller dimension, i.e. has some 0 eigenvalues I could estimate $\sigma^2$ (PSD of the noise term) for this given frequency.

Is there any way to obtain the signal $X$ from here?

I don't really know if what I'm doing is ok or not, and I would really appreciate if you could give me some insights.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.