This post describes how to apply reinforcement learning algorithm to trade Bitcoin. This repository provides an implementation aims to reproduce the result.

• BnH

  A buy-and-hold strategy that always hold 2 Bitcoins starting from the beginning of the test period.

• RL

  A trained RL agent making trading decisions to hold 0~4 Bitcoins given the current market condition.

• MMT

  A momentum strategy that holds 4 Bitcoins when the 30-period SMA cross-over than the current closing price and 0 Bitcoin otherwise.

Dependencies

• Python3.6
• NumPy 1.17.1
• Pandas 0.25.1
• Matplotlib 3.1.1
• PyTorch 1.2.0 (CPU only)

The minute-by-minute data is downloaded from Kaggle. I resample them into 15-minute interval and compute all the features we need. Then I save the two dataframes under bitcoin-historical-data.

Note that,

• I delete the row indexed 2017-04-15 23:00:00 after resampling since there is a clear error. This is done in the remove_outlier() method under the Data class.

• Due to request, I include the 15-minute data in bitcoin-historical-data (due to size constraint on GitHub, I cannot update the 1-minute data and the feature dataframe generated from the 15-minute data.)

# E.g. clone to local (say to Downloads)
cd ~/Downloads/trading-bitcoin-with-reinforcement-learning/

# Usage: python main.py <path-to-one-minute-data>
# If argument not provided, the default file path
# './bitcoin-historical-data/coinbaseUSD_1-min_data.csv' is given
python main.py ./bitcoin-historical-data/coinbaseUSD_1-min_data.csv

Note: I observed substantial variability in the test result therefore the equity curve you got may not be 100% the same as mine.