Open Foristkirito opened 4 years ago
@borisbanushev It seems when using ARIMA, Fourier transforms, VAE, there is possible that use future information, could you please tell me how you tried to avoid this ? Thanks !
I don't consider iFFT as a suitable way for prediction. And I wonder how it helps here?
My guess is, that he decomposes the whole time series with fft and arima and used the very same features as variables in the GAN/Xgboost.
should be alright, as long as you are not underfitting on anything. Everything up-until the GAN is just the "features" part of the dataset that the GAN trains on. Think of it like a very black-boxy-hypercomplex AIRMA. As long as you are cleaning up your data for extremes it should work about like "fine" as far as I can see. I think there is some substance here, just needs more ML research to robustifiy it for better hidden layer regressions (I think that's why he added bayesian opts).
I have similar concern.