Some different thoughts about SFM,1st using wavelet ,2nd using CNN as AEs to reduce complexties for RNN to learn

[emacsenli666]

great work and great architecture,but I think for ( financial ) time-series ,DFT or FFT is not enough,because such kind of transform does not provide time-frequency information at the same time . freqs can be obtained but for a certain length time-series ,DFT or FFT do not provide any infornation when are the main freqs happening or when vanished

so I have been thinking about using traditional way of signal processing ,wavelet-transform or EMD-HHT,which provide clear information of time and instantaneous freqs.but it's hard to combine signal processing procedures into a RNN ,may be 2-D LSTM will work better,1st-D for time information and 2nd-D for freqs .

second ,since CNN has really good achievement for classification ,so I guess add some CNNs as autoencoders might improve result

btw ,according to the experiments of the article , SFM have learned the capability of predicting time-series somehow 20 steps away ,but there are no information about how many steps been sent into SFM ,so can you give some hints?

[yakouyang]

I have the same confusion.  Generally，the format of the LSTM input is [batch, seq, features]. But here X_train received by model.fit has a shape (50,2014,1)，it seems to train with a complete sequences.I don't know if keras has the tricks to support this way.But my reproduce in pytorch failed.