Every trading algorithm has its flaws; however, I wanted to apply reinforcement learning to trading and see what results could be obtained. This model, a DDPG, slightly overfits the data, but some feature changes may solve this issue.  This bot was trained on BTC, LTC, ETH from 2017-04-01 until 2019-12-01.  This bot is overly aggressive, which may be due to the environment it was trained on. This trend may also be attributed to the algorithm’s tendency to overestimate the Q values of the critic (value) network. Currently, I am exploring better features and hoping I can solve the overestimating problem with either TD3 or SAC. Overall, I was extremely pleased with this bot’s potential. I hope this inspires you to learn more about AI and its vast applications!

Author