Creating my own Reinforcement Learning TradingAgent. I suceeded at making a profitable bot (with fees ~ 0.1% = Binance Trading Fees)
My projet needs both of my personal projets to work.
- The first my Gym Trading Environment (the simulation in which the agent will be trained)
- The second is my Rainbow Agent (a improved version of Deep Q Network aka. the AI)
Run manage_file.ipynb cells to download datasets and process them. (BTC/USDT and ETH/USDT from 3 exchanges)
Then, run training.ipynb cells to train the agent in the environment with to preprocessed datasets.
Validation data are fully seperated from training data (by date) to guarantee safe results. One validation is made every 30 000 steps on 5 parallelized training environment. Then the agent is evaluated on 5 validation environment.
After 1h of training :
Market Return : -36.93% | Portfolio Return : -4.01% | Position Changes : 12.43% | Max Drawdown : -37.56%
Market Return : -36.93% | Portfolio Return : -19.18% | Position Changes : 12.51% | Max Drawdown : -49.14%
Market Return : -24.95% | Portfolio Return : -14.29% | Position Changes : 11.30% | Max Drawdown : -29.71%
Market Return : -36.78% | Portfolio Return : 7.91% | Position Changes : 11.34% | Max Drawdown : -29.87%
Market Return : -24.96% | Portfolio Return : -3.79% | Position Changes : 11.69% | Max Drawdown : -30.22%After 2h of training :
Market Return : -36.99% | Portfolio Return : 114.42% | Position Changes : 8.90% | Max Drawdown : -31.45%
Market Return : -36.96% | Portfolio Return : 91.60% | Position Changes : 9.24% | Max Drawdown : -42.20%
Market Return : -36.96% | Portfolio Return : 98.06% | Position Changes : 9.25% | Max Drawdown : -42.20%
Market Return : -36.93% | Portfolio Return : 78.73% | Position Changes : 9.58% | Max Drawdown : -35.20%
Market Return : -36.94% | Portfolio Return : 127.78% | Position Changes : 9.30% | Max Drawdown : -37.20%Then, the agent begin to overfit.