search

ZhengyaoJiang / PGPortfolio

PGPortfolio: Policy Gradient Portfolio, the source code of "A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem"(https://arxiv.org/pdf/1706.10059.pdf).
GNU General Public License v3.0
1.74k stars 750 forks source link

using volume information #66

Closed ziofil closed 6 years ago

ziofil commented 6 years ago

Has anyone been experimenting with volume information? I've tried adding it as another feature (as if it was another layer in addition to close, high, low) and normalizing it to its last values, a bit like we do for the closing prices. It doesn't seem to help very much, perhaps I'm not doing it right?

trevorprater commented 6 years ago

I’ve also not had much luck using volume as a fourth feature.

On Fri, Mar 9, 2018 at 19:10 ziofil notifications@github.com wrote:

Has anyone been experimenting with volume information? I've tried adding it as another feature (as if it was another layer in addition to close, high, low) and normalizing it to the last values, a bit like we now do for the closing prices. It doesn't seem to help very much, perhaps I'm not doing it right?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/ZhengyaoJiang/PGPortfolio/issues/66, or mute the thread https://github.com/notifications/unsubscribe-auth/ABON8NhYpVg3wGg9FQhve5JlN7nDf3o6ks5tcxn5gaJpZM4SlCdB .

schmidtj commented 6 years ago

I have made similar changes and haven't had luck with gaining an advantage by using volume. I've changed the normalization to use the last value for each feature, this seems to work with the prices (close, high, low), but does not result in any benefit for volume. I thought that it might be due to the volume changing more drastically than prices, so I tried scaling and standardization, to no avail.

lytkarinskiy commented 6 years ago

I think that it's possible to use volume information if we have separate buy/sell volumes, in this case we can use buy/sell volumes ratio as bull/bear marker. But total volume doesn't tell us anything additional I think.

ZhengyaoJiang commented 6 years ago

@ziofil @trevorprater @schmidtj Thanks for sharing information about this. We also haven't found a way to properly incorporate volume information.

akaniklaus commented 6 years ago

@lytkarinskiy It should tell something additional actually. If volume is high and price drops, that would mean the sell volume is higher; vica, versa. Anybody tried having ADOSC (Chaikin A/D Oscillator) or MFI (Money Flow Index) indicators of TA-Lib, which uses volume as an input as features? Did anybody also tried multi-resolution (multi-period) features in the input vector? @ziofil @trevorprater @schmidtj That could prevent the agent from suffering pump & dump schemes which is a big problem occuring in low-cap assets.