search

tcloaa / Deep-Portfolio-Theory

Autoencoder framework for portfolio selection (paper published by J. B. Heaton, N. G. Polson, J. H. Witte.)
128 stars 63 forks source link

'autoencoder.h5' #3

Closed rchaaben closed 6 years ago

rchaaben commented 6 years ago

Hi, may I please know which encoder are you using because it seems that you are using a predefined encoder 'autoencoder.h5' and there is no information about it (I'm not able to run the code). Best regards.

tcloaa commented 6 years ago

My bad..

'autoencoder.h5' is a trained model that I used when I was doing this research. I didn't include it in this repo for now... (sorry I was just busy that I don't have time to take care of it right now).

You can build one yourself starting from here.

I will upload it very soon, but sorry again for the inconvenience...

skyhoudou commented 6 years ago

Hi, i am also very interested in this paper. And i meet across a problem. Do I need to normalize the stock_net_train before training the auto-encoder?

On Tue, Nov 28, 2017 at 5:52 AM, tcloaa notifications@github.com wrote:

Closed #3 https://github.com/tcloaa/Deep-Portfolio-Theory/issues/3.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tcloaa/Deep-Portfolio-Theory/issues/3#event-1360631497, or mute the thread https://github.com/notifications/unsubscribe-auth/ATFZet1VJ6aZFMViFlzpLcNnilfO88Wtks5s65GUgaJpZM4QoCUS .

tcloaa commented 6 years ago

Good question.

The paper does not mention about the data preprocessing, even the type of data (weekly return of what?: close price? percentage of change? net change?)

So I chose percentage of change and did not do any normalization. What do you think?

skyhoudou commented 6 years ago

exactly, i tried to train the percentage net change, but the result is not good. Especially when I try to track the ibb index, the performance of the predicted ibb index of selected stocks s25, s45 and s65 are bad. I am trying to figure out if the parameters of neural networks are not suitable. By the way, where did you find this data set? i want to collect more data to see if the performance would be improved by enlarge the size of data set.

On Tue, Nov 28, 2017 at 1:26 PM, tcloaa notifications@github.com wrote:

Reopened #3 https://github.com/tcloaa/Deep-Portfolio-Theory/issues/3.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/tcloaa/Deep-Portfolio-Theory/issues/3#event-1361249393, or mute the thread https://github.com/notifications/unsubscribe-auth/ATFZei2reDxhkTnovcka8DwtVY4_xlzyks5s6_vdgaJpZM4QoCUS .

tcloaa commented 6 years ago

@skyhoudou Excel + Bloomberg add-in, using VBA Macro.