Serving the Quantitative Finance Community

 
User avatar
tagoma
Topic Author
Posts: 18583
Joined: February 21st, 2010, 12:58 pm

Invalid actions in RL-based trading setups

August 22nd, 2020, 9:35 pm

Hi everyone.
Can you please kindly suggest resources on the so-called invalid actions more specifically in RL-based trading setups so that to account for the quite pervasive contraints/limits in real-life trading/portfolio management (risk limits, capital, ...). Maybe you guys can also share bits of your own experience and point to solutions/guidelines/best-practices?
Thank you so much.
 
User avatar
tagoma
Topic Author
Posts: 18583
Joined: February 21st, 2010, 12:58 pm

Re: Invalid actions in RL-based trading setups

September 24th, 2020, 7:56 pm

FWIW I'm still working on it. The practical solutions I came through pretty much resolve around masking parts of the environment at specific steps (e,g, there are some bits in google's tf-agents code), hence involving the observation space, while I would rather need the possibility to constraint the action space at a given step independently of the observation space. I'm not sure it makes sense from the pure RL perspective, I admit.