This project provides a general environment for stock market trading simulation using OpenAI Gym. Training data is a close price of each day, which is downloaded from Google Finance, but you can apply any data if you want. Also, it contains simple Deep Q-learning and Policy Gradient from Karpathy's post.
In fact, the purpose of this project is not only providing a best RL solution for stock trading, but also building a general open environment for further research.
So, please, manipulate the model architecture and features to get your own better solution.
- Python2.7 or higher
- Numpy
- HDF5
- Keras with Beckend (Theano or/and Tensorflow)
- OpenAI Gym
Note that the most sample training data in this repo is Korean stock. You may need to re-download your own training data to fit your purpose.
After meet those requirements in above, you can begin the training both algorithms, Deep Q-learning and Policy Gradient.
Train Deep Q-learning:
$ python market_dqn.py <list filename> [model filename]
Train Policy Gradient:
$ python market_pg.py <list filename> [model filename]
For example, you can do like this:
$ python market_pg.py ./kospi_10.csv pg.h5
Aware that the provided neural network architecture in this repo is too small to learn.
So, it may under-fitting if you try to learn every stock data.
It just fitted for 10 to 100 stock data for a few years. (I checked!!)
Thus you need to re-design your own architecture and
let me know if you have better one!
Below is training curve for Top-10 KOSPI stock datas for 4 years using Policy Gradient.
- Test environment to check overfitting.
- Elaborate the PG's train interface.
[1] Playing Atari with Deep Reinforcement Learning
[2] Deep Reinforcement Learning: Pong from Pixels
[3] KEras Reinforcement Learning gYM agents, KeRLym
[4] Keras plays catch, a single file Reinforcement Learning example