0
$\begingroup$

I'm working on a NN that uses Wavlet Transformed signals (with different wavelets and levels) and combines them with an additional Statistical Features input (input_4) to provide one step ahead predictions of the given 1D signal,but i have encountered a problem, When transforming the signal all at once the model seems to perform adequately, however as i back test it by feeding the Model one value at a time and doing all the processing on a step by step basis, the models performance drops to a random guess,(Presumably due to the boundary effect of the WT).

My question is what can i do to avoid this problem, and are Wavelets even suitable in a setting where the data gets updated frequently i.e Time-Series.

wden(x,'sqtwolog','soft','mln',2,'db2')

I am using a python implementation of the wden function to decompose and reconstruct the signal in many different combinations of wavelets and levels with the PyWavelets package.

P.S A rolling window Wavelet Transform significantly degrades the models performance to random guess as well.

Model Structure

$\endgroup$
  • $\begingroup$ So previous research on the topic suggests that this could also be an issue of using DWT which is sensitive to shifts in the data, a possible solution could be to use SWT but haven't tried that yet. $\endgroup$ – jetychill Mar 22 at 16:08

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.