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

model response to coins trained on periods prior to their existance. #18

Closed AhmMontasser closed 6 years ago

AhmMontasser commented 6 years ago

Hello,

I don't understand how the model response when a selected coin is not existing at the training period.

for example, choosing start date = 2015-07-01 and end date = 2017-11-01, BCH is chosen The training/ backtest portfolio value decreases tremendously. is that because of the fact that BCH didn't exist most of that time ?

ZhengyaoJiang commented 6 years ago

The missing data in training set would be padded as 1.

AhmMontasser commented 6 years ago

Thank you for your reply, do you think training on 1s would affect the model performance? does it worth coding a selection system for training to make sure coins exists over the whole training period ?

ZhengyaoJiang commented 6 years ago

Yes, it is possible that 1s would affect the model performance but it is also possible that not taking new coins into consideration would let the agent loss chance to gain profit. I'm not sure which is the real case. You can try to do an experiment if you are interested in this topic.