I have many timeseries of varying crest and troughs amplitudes, and I want to get features out of it from a window of size 50.
I am sampling a continuous batch of 500 timesteps from the timeseries and training them.
Since all the series have different amplitudes, to avoid scaling I am using percentage change to scale the data, and then taking window of size 50.
Now, is it ok to apply nn.conv1d on this percentage window to get features from the window.
Is there a better way I can features out of window without using scaling.
Below is my current implementation.
data is a dataframe
data = data[['Value']] data['ret'] = data['Value'].pct_change() data = data.dropna().reset_index()[['Value','ret']] data = torch.FloatTensor(data.values)
and now I take window of size 50 , and calculate the reward based on ‘Value’ , but pass ‘ret’ in Neural network to get features.