search

notadamking / RLTrader

A cryptocurrency trading environment using deep reinforcement learning and OpenAI's gym
https://discord.gg/ZZ7BGWh
GNU General Public License v3.0
1.73k stars 540 forks source link

Sortino agent overtrading #56

Closed TheExGenesis closed 5 years ago

TheExGenesis commented 5 years ago

The blog post suggests the Sortino reward function avoids over-trading, but on both the original and the fixed date dataset I see wild trading rates close to 50% that invariably result in a loss of more than to 50% of the starting budget.

This is by running train.py right out of git clone. Not sure if a bug or if I'm missing something.

TheExGenesis commented 5 years ago

net-worth-history-test Full training + test set run

rshtirmer commented 5 years ago

Main.py is not in this repo.. are you sure you are cloning the correct repo?

Sent with GitHawk

TheExGenesis commented 5 years ago

I meant train.py

Ray-0403 commented 5 years ago

How about u just modify the action space to let the agent hold only one position each time and train it?

TheExGenesis commented 5 years ago

Thanks, I'll try that. But since this is code straight from the repo, is it possible that it's a bug? Since others have trained from scratch and achieved good results, could this be a matter of versioning of the libs?

TheExGenesis commented 5 years ago

Just got it to work, after yet another clean clone. Close.