The repository contains implementation of an attempt to train RL agent to trade in the financial market using a Deep Direct Reinforcement Learning model through Technical Analysis methodon the data to provide more robust framework on top of the data using minute-based HKEX market data. Refer to the UROP_1100_Report.pdf for detailed explanation.
The RL agent and its environment is defined in the ./model/ directory, the data is in ./data/ directory, and scripts for pre-processing the data is present in the ./data_wrangling/ directory.
Detailed documentation is provided on README.md in each folder and its corresponding files.
To use and train the model, refer to the 3 notebooks where the model is trained on 3 different dataset. Refer to the DRL_BOC+CNOOC.ipynb for the most updated and refactored code with its documentation.
To use the your own data, replace and edit the data in the ./data/ directory.
Using conda package manager, build the environment with environment.yml:
$ conda env create -f environment.ymlOr without conda:
- Python 3.6
- TA-Lib >= 0.4.17. On some machines, underlying
TA-Liblibrary should be installed in advance before installing using pip or other package manager. Refer here for the predecessor library. - Pytorch >= 0.4.1
- tqdm >= 4.28.1
- pyfolio >= 0.9.0
- seaborn
- numpy
- pandas
- jupyter notebook
- and other basic packages.
Clone the repo, and build virtual environment using conda for easier installation through the using the step above.
Important note: this project utilises TA-Lib library for python (Technical Analysis) as a part of the knowledge discovery on the data. It requires the predecessor TA-Lib library to be installed on your machine before the TA-Lib library for python can be installed properly (i.e. using pip install talib or conda).
To install predecessor TA-Lib, refer here.
- (https://github.com/yuriak/RLQuant/)
- Recurrent Reinforcement Learning: A Hybrid Approach
- Deep Direct Reinforcement Learning for Financial Signal Representation and Trading
- Using a Financial Training Criterion Rather than a Prediction Criterion
- A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem
