https://github.com/kh-kim/stock_market_reinforcement_learning

This project provides a stock market environment using OpenGym with Deep Q-learning and Policy Gradient.
https://github.com/kh-kim/stock_market_reinforcement_learning

Last synced: 3 months ago
JSON representation

This project provides a stock market environment using OpenGym with Deep Q-learning and Policy Gradient.

• Host: GitHub
• URL: https://github.com/kh-kim/stock_market_reinforcement_learning
• Owner: kh-kim
• Created: 2016-10-04T14:42:19.000Z (over 8 years ago)
• Default Branch: master
• Last Pushed: 2016-12-23T07:34:59.000Z (about 8 years ago)
• Last Synced: 2024-08-01T10:21:54.132Z (6 months ago)
• Language: Python
• Size: 368 KB
• Stars: 787
• Watchers: 75
• Forks: 323
• Open Issues: 18
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-ai-in-finance - stock_market_reinforcement_learning - Stock market trading OpenAI Gym environment with Deep Reinforcement Learning using Keras. (Strategies & Research / Time Series Data)

README

        
# Stock Trading Market OpenAI Gym Environment with Deep Reinforcement Learning using Keras

## Overview

This project provides a general environment for stock market trading simulation using [OpenAI Gym](https://gym.openai.com/). 

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](http://karpathy.github.io/2016/05/31/rl/).

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.**

## Requirements

- Python2.7 or higher

- Numpy

- HDF5

- Keras with Beckend (Theano or/and Tensorflow)

- OpenAI Gym

## Usage

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  [model filename]

Train Policy Gradient:

	$ python market_pg.py  [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.  

![Training Curve](./pg_over_top_10.png)

## To do

- Test environment to check overfitting.

- Elaborate the PG's train interface.

## Reference

[1] [Playing Atari with Deep Reinforcement Learning](http://arxiv.org/abs/1312.5602)  

[2] [Deep Reinforcement Learning: Pong from Pixels](http://karpathy.github.io/2016/05/31/rl/)  

[3] [KEras Reinforcement Learning gYM agents, KeRLym](https://github.com/osh/kerlym)  

[4] [Keras plays catch, a single file Reinforcement Learning example](http://edersantana.github.io/articles/keras_rl/)