Deep Trading Agent
Deep Reinforcement Learning based Trading Agent for Bitcoin using DeepSense Network for Q function approximation.
For complete details of the dataset, preprocessing, network architecture and implementation, refer to the Wiki of this repository.
- Python 2.7
- Pandas (for pre-processing Bitcoin Price Series)
- tqdm (for displaying progress of training)
To setup a ubuntu virtual machine with all the dependencies to run the code, refer to .
Run with Docker
Pull the prebuilt docker image directly from docker hub and run it as
Build the docker image locally by executing the command and the run the image as
docker run -p 6006:6006 -it deep-trading-agent
This will setup the repository for training the agent and
mount the current directory into in the container
during image build, the latest transactions history from the exchange is pulled and sampled to create per-minute scale dataset of Bitcoin prices.
Significant Trades Tool
This dataset is placed at
to initiate training of the agent, specify suitable parameters in a config file (an example config file is provided at ) and run the code using
training supports logging and monitoring through Tensorboard
and are installed in the container to edit the configuration files and run
bind port 6006 of container to 6006 of host machine to monitor training using Tensorboard
Please give a to this repository to support the project .
Improve Model performance
is inspired by Deep Q-Trading where they solve a simplified trading problem for a single asset.
For each trading unit, only one of the three actions: neutral(1), long(2) and short(3) are allowed and a reward is obtained depending upon the current position of agent. Deep Q-Learning agent is trained to maximize the total accumulated rewards.
Current Deep Q-Trading model is modified by using the Deep Sense architecture for Q function approximation.
Per minute Bitcoin series is obtained by modifying the procedure mentioned in this repository.
Transactions in the Coinbase exchange are sampled to generate the Bitcoin price series.
Refer to to download the dataset.
Completely ignore missing values and remove them from the dataset and accumulate blocks of continuous values using the timestamps of the prices.
All the accumulated blocks with number of timestamps lesser than the combined history length of the state and horizon of the agent are then filtered out since they cannot be used for training of the agent.
In the current implementation, past 3 hours (180 minutes) of per minute Bitcoin prices are used to generate the representation of the current state of the agent.
With the existing dataset (at the time of writing), following are the logs generated while preprocessing the dataset:
Process missing values and concatenate smaller blocks to increase the sizes of continuous price blocks.
Standard technique in literature to fill the missing values in a way that does not much affect the performance of the model is using exponential filling with no decay.
(To be implemented)
Tensorflow "1.1.0" version is used for the implementation of the Deep Sense network.
Implementation is adapted from this Github repository with a few simplifications in the network architecture to incorporate learning over a single time series of the Bitcoin data.
Deep Q Trading
Implementation and preprocessing is inspired from this Medium post.
The actual implementation of the Deep Q Network is adapted from DQN-tensorflow.