Stationary input features

Course Name: Deep Reinforcement Learning in Trading, Section No: 10, Unit No: 5, Unit type: Video



Hi,  I have some questions/observations on the content of this unit:



1) why do ANNs need stationary input features? I mean, in order to estimate time series model and to make inference you need stationary processes but do they need stationary inputs in general?



2) NNs need normalized input features, you can use z-scores to normalize data among other methods



3) taking z-scores DO NOT make a time series stationary (differentiating, either fractional or unit, does).



4) In order to normalize a time series you need to have stationary processes or use rolling windows (only a proxy). Usually, take z-scores on differentiated time series



What am I missing?



Thank you,



Jacopo

Hi Jacopo,



Thanks for your feedback, we really appreciate your thoughts.

Regarding point (1) your are correct that inputs don't need to be strictly stationary. However, we found that this constraint helps to produce significantly higher quality trades.

For the other points, I agree with your observations and we will make sure that the content is updated asap.

Again, thanks so much for helping us to improve this course.



Best regards,



Tom

Thanks to you Tom and all the Quantra team, really appreciating the course and all the useful material!

Hi Jacopo!



Thanks for your feedback.

The content has been updated accordingly!



Warm regards,

Gaurav