gym-crypto 0.0.3

Reinforcement Learning OpenAI Gym environment for crypto-currency trading.

gym-crypto

OpenAI Gym compatible environment for crypto-currency trading. The environment allows to change the currency the bot trades, the granularity of trading and starting capital of the agent. More configurability to come in the future.

Observation Space

The observation space is a tuple structured as follows:

(OCHLV, Current Capital, Weight of the capital invested)

OCHLV represents Open, Close, High, Low prices and Volume respectively.

Action Space

Action space is a single scalar between -1 and 1, which represents the percentage of the capital that the agent has invested into the currency at the last time step.

-1 representing a full short position and 1 representing a full long position.

Rewards

The reward is defined as the total amount of capital that has been gained/lost for that time-step. This number includes the return on the investment as well as the transaction costs.

Usage

In order to train the agent, the enviroment first has to be configured. This can be done as follows:

import gym

env = gym.make("gym-crypto-v0")

start_date = '2018-01-01'
end_date = '2019-01-01'

currency = 'BTC'
granularity = '15min'
transaction_pct = 0.001
capital = 100

kwargs = {'currency':currency, 'granularity':granularity,
          'transaction_pct':transaction_pct, 'capital':capital,
          'start_date': start_date, 'end_date':end_date}

env.configure_env(**kwargs)

To get some results from the agent, a env.render() needs to be called, which will return the capital returns as the first variable and weight allocation as the second variable. More detailed results will be coming in the future.

Data

To futher inspect the data two methods can be called from the data_utils module.

from gym_crypto.data_utils.data_utils import *

currencies = get_avaliable_currencies()
full_data = read_data('BTC')

All data was obtained from the Bitfinex API.