Welcome to FiRe! FiRe is a Reinforcement Learning (RL) pipeline built to research Single- and Multi-Objective trading algorithms.

• Critic-only Deep Q-Learning RL agent with Hindsight Experience Replay link
• Single- and Multi-Objective reward learning with generalization in the sense of [Friedman, Fontaine]
• MLP-type agent’s Neural Network
• Four reward mechanisms (last logarithmic return, average logarithmic return, Sharpe Ratio, profit only when closing)

We recommend python version 3.10. (Check your version with python3 --version)

We recommend using a virtual environment: to create it, run

python3 -m venv .venv
source .venv/bin/activate

in the repository's main folder.

Install the required packages and download the Bitcoin dataset by running

make

in the main folder.

To test the code you can run the sin-wave example:

python3 main.py parameters/sin_wave_test_multi.json

In order to run a particular experiments run the main.py script adding as argument a json configuration file:

python3 main.py <your configuration file>.json

the json configuration file contains information about:

• The chosen dataset
• The reward or set of rewards to use
• Q-Learning general parameters
• Neural Network hyperparameters

Experiments results are stored in the results_of_experiments folder, each subfolder contains the results of each single experiment.

We provide all configuration files for the experiments on the Bitcoin dataset that are shown in the paper. To run the experiments use:

git checkout e2bc651b07eec388ee0b03228e0acdee8550ba7d
sh scripts/run_experiments_in_folder.sh parameters/paper_plots

For the simulation we discussed in our paper we used the following datasets:

• Cryptocurrency price data (BTCUSDT, ETHUSDT, XRPUSDT) from Binance exchange, downloaded through Binance API
• Stocks price data (AAPL, SPY), downloaded from Yahoo!finance through yfinance python package
• Nifty50 minute close price, downloaded from Kaggle

You can download all cryptocurrencies and stocks datasets by running:

make datasets

that will run the download_datasets.sh script (this is not necessary if you already ran make)

Nifty50 dataset can instead be downloaded from kaggle at this link

If you want to use latex to create the plots you need to install the following packages:

sudo apt-get install texlive-latex-extra texlive-fonts-recommended texlive-fonts-extra dvipng

You can now set "text.usetex": True, at the top of src/plots.py.

Our code utilizes parts of the following open source repositories:

1. The Deep Q-Network implementation at https://github.com/mswang12/minDQN
2. The stocks environment at https://github.com/AminHP/gym-anytrading