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
1.73k stars 748 forks source link

Explanation of net_config.json #79

Open andresespinosapc opened 6 years ago

andresespinosapc commented 6 years ago

Can you please update the User Guide explaining the meaning of the net_config.json options?

wujun commented 6 years ago

most of the settings are about to neural network. you need have knowledgebase for deep learning.

jsztompka commented 6 years ago

Could someone post the recommended settings file which give a good results? I'm under impression the checked in config is more of an example (batch size and learning rate appear to be arbitrary numbers).

ErnstDinkelmann commented 6 years ago

@jsztompka Here is the json that achieves performance in line with the paper. It would look slightly different as I've just ordered things as it makes sense to me. I've also included some defaults explicitly as I wanted all the parameters in one place.

net_config.zip

It's very important to realize that these parameters are quite specific to the period in question. A different period would require different parameters - that's where hyper parameter optimisation comes in. This I think is part of the Author's private work though.