dancingonmoon / financerobot_rl Goto Github PK

the project is to apply DQN, DDQN and PPO upon BTC trading. it accomplishes below tasks:

• download OHLC data from encryption exchange house;
• generate features data via OHLC data, which include:
  • volume
  • RSI
  • Log_close_weeklag
  • Log_high_low
  • Log_open_weeklag
  • open_pre_close
  • high_pre_close
  • low_pre_close
  • num_trades
  • bid_volume
  • horizon_price
  • close
• data normalization with previous 365 lookback
• set up environment, with the class Name: Finance_Environment_V2
  • action_space: long and short are all applied during the course of training, while in the course of backtest, only short is not advised.
    • buy = 2
    • hold = 1
    • sell = 0
  • reward function:
    • reward1: positive return makes reward1 = 1 , while negative return makes reward1 = -1; reward1 = 0, when action is on hold;
    • reward2: np.log(margin_ratio) * (action - 1)
    • reward = reward1 + reward2 * 5
• Training either by DQN , or DDQN , or PPO, and save the trained_model
• Backtesting:
  • vector Backtesting;
  • event based Backtesting: short is not recommended;
  • to produce plotly chart