This repository is inspired from: https://github.com/pythonlessons/RL-Bitcoin-trading-bot
Basically a random stock trader, trained actor-critic stock trader and augmented data trained actor-crtic stock trader are created and compared.
Instead of constant gamma in learning of actor-critic, I modified into omega function which gives larger weight to small reward in current action, and small weight to large reward in current action, so that it can promote further look a heads for sequence of states with current bad action.
For data augmentation, I used the Fourier phase and amplitude spectrum as being generated by a GAN architecture, which are then recombined in time domain to generate openning and closing prices as my augmented environment.
These models were created with no fine tuning and results are compared only for 100 episodes. Following table shows the net worth of the agent after its actions of buring, selling or holding on given environment of openning and closing prices.
