AlexanderGuillermoSeguraBallesteros / RRL

Recurrent Reinforcement Learning (RRL)This is a repository for the implementations of RRL, mainly following Moody's work, other authors will be given credit as well on the go.
9 stars 4 forks source link

Further Use of this work #1

Open hddy2000 opened 6 years ago

hddy2000 commented 6 years ago

Dear Alexander, I'm doing my MF research recently, very interested in this RL by Moody,do you know if there's any further improvement on this model? Did you do any further modification on it? Best, Eddie

AlexanderGuillermoSeguraBallesteros commented 6 years ago

Hello, it's great to hear of someone interested on this area of research. The model has existed for a while now, so there have been many improvements. To my knowledge, the most important has been the use of Deep RL. You can start from this work,

Deep_Direct_Reinforcement_Learning_for_Financial_Signal_Representation_and_Trading.pdf

For the most part is the same work as Moody's, however the inputs layer is not taken as raw data from the time series, but replaced by a Deep Network with a Fuzzy layer in between.

May I ask what is your research purpose? Are you currently involved in some research program? Or are you working on your free time.

hddy2000 commented 6 years ago

Alexander, Thank you for your response! I am an AI engineer interested in a career in Mathematical Finance, thus I'm doing this research on my free time:)

I also found some works done by Professor Dempster from Cambridge University: http://www.virtual.jbs.cam.ac.uk/fileadmin/user_upload/research/workingpapers/wp0030.pdf https://www.jbs.cam.ac.uk/fileadmin/user_upload/research/workingpapers/wp0306.pdf

It seems that there are indeed a lot of explorations since Moody's time. BTW Have you ever seen any implementation of this work in live trading? Did any of this work out or it is just ideally working? Best, Eddie

AlexanderGuillermoSeguraBallesteros commented 6 years ago

Hello, sorry for not replying earlier, unfortunately I have not seen this implementations in real trading, by no means I am saying these are not used, but most of the real implementations are not shared you know, because there is some economic gain associated. Moreover, most trading companies need to keep their breakthroughs as secret in order to have an alpha (advantage), keep in mind this is a zero sum game, and if these strategies where of public knowledge, they would become useless, which is by the ways one of the considerations when doing back testing. I however am interested in testing these in real life, I don't have a brokers account so the best shot is with quantopian or quantiacs.