Serving the Quantitative Finance Community

 
User avatar
MartinGale7
Topic Author
Posts: 70
Joined: April 1st, 2011, 7:42 am

Re: PCA, or an alternative orthoganalisation?

December 27th, 2017, 8:02 am

Great. Did you find that your LSTM/RNN worked better this task than NNs without temportal nodes, get Backprop?

If I wanted to tinker with this, any libraries you could recommend for C#? I'm already using Accord.Net which has simple nets in it, but I suspect it may be a little old.
 
User avatar
outrun
Posts: 4573
Joined: January 1st, 1970, 12:00 am

Re: PCA, or an alternative orthoganalisation?

December 27th, 2017, 3:45 pm

Yes, lstm improve performance. Using NN in general is easier when you have a lot of data, working with daily data requires a lot of attention to prevent overtraining.

I'm not using C#, and you won't find much tutorials or example code. I think it's worthwhile to invest in getting jupyter up and running: python + scipy + tensorflow in a (local) webbased environment. Thst way you can connect with the largest community and the code they publish. The easiest way to set it up is via docker. I can send some slides with the steps on how to set that up? I'm still on holidays but I have them on my phone :-). Besides that it would still be nice to exchange emails addresses via pm?
 
User avatar
ISayMoo
Posts: 2332
Joined: September 30th, 2015, 8:30 pm

Re: PCA, or an alternative orthoganalisation?

December 27th, 2017, 5:06 pm

 
User avatar
katastrofa
Posts: 7440
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: PCA, or an alternative orthoganalisation?

December 27th, 2017, 8:11 pm

Using PCA for forecasting sounds insane to me. How can one select the regressor without even looking at Y? So what that PCA decides that the main modes of the covariance matrix of X are x1, x2 and x3, while x8 and x9 are the least important? Y may depend only on x8 and x9.

Anyway, I usually do a multivariate regression with factors and their number selected by e.g. BIC.
 
User avatar
ISayMoo
Posts: 2332
Joined: September 30th, 2015, 8:30 pm

Re: PCA, or an alternative orthoganalisation?

December 27th, 2017, 9:08 pm

Loewdin orthogonalisation can be useful to give more stability to multivariate regression: https://www.wou.edu/~beavers/Talks/LowdinOrth.pdf
 
User avatar
MartinGale7
Topic Author
Posts: 70
Joined: April 1st, 2011, 7:42 am

Re: PCA, or an alternative orthoganalisation?

December 28th, 2017, 7:48 am

@ISayMoo Thank you for both of the doucments. I am reading these now.

@katastrofa I totally agree on the insanity. That's the reason for the post ;) Obviously the PCA is only on the inputs, BEFORE using a machine learning algorithm to train those inputs to a target (totally unrelated to the PCA process). But, you have highlighted the issue that if x8 and x9 have the greatest importance then PCA will allow them to be diluted. Many of my inputs are correlated, hence the need to reduce them with PCA first. When you say 'multivariate regression with factors', do you mean linear (ie a n dimensional plane), as opposed to polynomical or GPR etc? I have used BIC and AIC in the past and they are useful, esp when adding a parameter (or input) at a time.