Awesome

TradzQAI

Trading environnement for RL agents, backtesting and training.

Live session with coinbasepro-python is finaly arrived !

Available sessions:
- Local
  - 1M bar datasets any size.
    - header should be ['Time', 'Open', 'High', 'Low', 'Close', 'Volume'] with ';' as separator
  - Tick datasets any size.
    - header should be ['Time', 'BID', 'ASK', 'VOL'] or ['Time', 'Price', 'Volume'] with ',' as separator
- Live
  - Gdax API
Available agents:
- DDPG
- DQFD
- DQN
- DQNN
- NAF
- PPO
- TRPO
- VPG
Available contract type:
- CFD
- Classic

TradzQAI has been inspired by q-trader.

Indicators lib come from pyti

More datasets available here

Status

Alpha in development
    GUI rework on track

Getting Started

Dependencies :
- Tensorflow
- Tensorforce
- coinbasepro-python
- Pandas
- Numpy
- tqdm
- h5py
Installation :

pip install -r requirements.txt

Running the project

Usage:
  python run.py -h (Show usage)
  python run.py -b agent_name (to manually build config file, it build config files from agent, default PPO)
  python run.py -s live (for live session) 
  python run.py -m eval (for eval mode)
  python run.py -c config_dir/ # load config from directory, make sure you have agent, env and network json files in it
  python run.py (Run as default)

When you run it for the first time, a config directory is created, you can change it to changes environnement settings and some agents settings. It save settings (env, agent, network) in a save directory, and create a new directory if make any changes.

Do you own decision function for maker side. For more info look at this function

from core import Local_session as Session
from mymodule import myfunc
session = Session(mode=args.mode, config=args.config)
session.initApi(key=key, b64=b64, passphrase=passphrase, url=url,
        product_id=product_id)
session.getApi().setBestPriceFunc(myfunc)

Do your own runner.

from core import Local_session as Session
session = Session() # Run with default values
session.loadSession() # loading environnement, worker and agent
session.start() # Start the session thread

Do your own worker.

from core import Local_env
env = Local_env() # run with default values
for e in episode:
  state = env.reset()
  for s in step:
    action = agent.act(state)
    next_state, terminal, reward = env.execute(action)
    agent.observe(reward, terminal)
    if terminal or env.stop:
      break
  if env.stop or e == episode - 1:
    env.logger._running = False #Close the logger thread
    break

How to use networks.

You have to define your input to fit with columns name of your datasets, it will automaticaly it grab input from network and compare it with your dataset columns in getState function, it allow you to do complex network like this :

[
    [
        {"names": ["Price"], "type": "input"},
        {"activation": "relu", "size": 8, "type": "dense"},
        {"activation": "relu", "size": 8, "type": "dense"},
        {"name": "pricenet", "type": "output"}
    ],
    [
        {"names": ["Volume"], "type": "input"},
        {"activation": "relu", "size": 8, "type": "dense"},
        {"activation": "relu", "size": 8, "type": "dense"},
        {"name": "volnet", "type": "output"}
    ],
    [
        {"names": ["pricenet", "volnet"], "type": "input"},
        {"activation": "relu", "size": 64, "type": "dense"},
        {"activation": "relu", "size": 32, "type": "dense"},
        {"activation": "relu", "size": 8, "type": "dense"},
        {"name": "prediction", "type": "output"}
    ]
]

Simple network are handled as well without defining any input:

  [
      {"activation": "relu", "size": 64, "type": "dense"},
      {"activation": "relu", "size": 64, "type": "dense"}
  ]

Also TradzQAI support pre trained keras model:
- You can build settings for your model by using py run.py -b DEEP. Your model have to be placed in the same directory as the one you use to launch it and have to be called deep_model.h5.

Relevant project

This project isn't perfect so keep this in mind.